Rebuild: dual-agent live demo with filesystem browser

Two AI agents (Qwen2.5-3B + SmolLM2-1.7B) collaborate in real-time
with a live filesystem tree showing every file created. Autopilot mode,
inline annotations, raw file viewer. Updated embedded core to v1.0.22.

README.md

@@ -1,5 +1,5 @@
 ---
-title: AgentAZAll - Persistent Memory & Multi-Agent Communication
+title: AgentAZAll - Dual-Agent Live Demo
 emoji: "\U0001F9E0"
 colorFrom: indigo
 colorTo: purple
@@ -8,7 +8,7 @@ sdk_version: "5.23.0"
 python_version: "3.12"
 app_file: app.py
 pinned: false
-short_description: "Three transports (AgentTalk, Email, FTP), one interface"
+short_description: "Watch two AI agents collaborate via filesystem in real-time"
 tags:
   - agent
   - memory
@@ -16,25 +16,29 @@ tags:
   - persistent-memory
   - tool-use
   - agenttalk
-  - email
-  - ftp
-  - smtp
-  - rest-api
+  - autopilot
+  - filesystem
 models:
+  - Qwen/Qwen2.5-3B-Instruct
   - HuggingFaceTB/SmolLM2-1.7B-Instruct
 preload_from_hub:
+  - Qwen/Qwen2.5-3B-Instruct
   - HuggingFaceTB/SmolLM2-1.7B-Instruct
 ---
 
-# AgentAZAll — Persistent Memory & Multi-Agent Communication
+# AgentAZAll — Dual-Agent Live Demo
 
-Three transports (**AgentTalk** · **Email** · **FTP**), one interface. Chat with an AI
-agent that actually **remembers** — and communicates with other agents over your choice
-of transport layer. This demo uses
-[SmolLM2-1.7B-Instruct](https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct)
-on ZeroGPU, powered by [AgentAZAll](https://github.com/cronos3k/AgentAZAll).
+Two AI agents (**Qwen2.5-3B** and **SmolLM2-1.7B**) collaborate in real-time using
+[AgentAZAll](https://github.com/cronos3k/AgentAZAll). The center panel shows the
+**raw filesystem** — watch actual files appear as agents send messages, store memories,
+and build shared knowledge. Everything is plain text. No database. No vector store.
 
-No proprietary APIs. No vendor lock-in. Open, self-hostable, interchangeable.
+## What You'll See
+
+- **Left panel**: Agent Alpha (Research Director) — asks probing questions, synthesizes insights
+- **Center panel**: Live filesystem browser showing every file the agents create
+- **Right panel**: Agent Beta (Creative Developer) — proposes implementations, explores ideas
+- **Autopilot button**: Watch both agents have a multi-turn conversation automatically
 
 ## Three Transports, One Interface
 
@@ -46,41 +50,18 @@ No proprietary APIs. No vendor lock-in. Open, self-hostable, interchangeable.
 
 Agents don't care which transport delivers their messages — switch by changing one line in `config.json`.
 
-## What You Can Do
-
-- **Chat** with an agent that stores and recalls memories across messages
-- **Send messages** between agents in a simulated multi-agent network
-- **Browse** the agent dashboard to see memories, inbox, and identity
-- **Watch** the agent use tools in real time (remember, recall, send, inbox)
-
-## Key Features
-
-- **Persistent memory** (`remember` / `recall`) that survives context resets
-- **AgentTalk transport** — modern HTTPS REST API; self-host or use the free public relay
-- **Email transport** (SMTP/IMAP/POP3) — universal agent communication
-- **FTP transport** — file sync over the original internet file protocol
-- **Identity continuity** (`whoami` / `doing`) across sessions
-- **Zero-dependency core** — Python stdlib only
-- **Unlimited local server** — self-hosted AgentTalk has no file size or message limits
-
 ## Quick Start
 
 ```bash
-# Public relay (zero config):
 pip install agentazall
 agentazall register --agent myagent
-
-# Self-host everything:
-agentazall setup --agent my-agent@localhost
-agentazall server --agenttalk     # HTTPS API (port 8484)
-agentazall server --all           # all three transports
 ```
 
 ## Links
 
-- **Install**: `pip install agentazall` — [pypi.org/project/agentazall](https://pypi.org/project/agentazall/)
+- **Project**: [agentazall.ai](https://agentazall.ai) — research paper, documentation
 - **Source**: [github.com/cronos3k/AgentAZAll](https://github.com/cronos3k/AgentAZAll)
-- **This Demo**: [huggingface.co/spaces/cronos3k/AgentAZAll](https://huggingface.co/spaces/cronos3k/AgentAZAll)
+- **Install**: `pip install agentazall` — [pypi.org/project/agentazall](https://pypi.org/project/agentazall/)
 
 ## License
 

app.py

@@ -1,18 +1,25 @@
-"""AgentAZAll HuggingFace Spaces Demo.
+"""AgentAZAll HuggingFace Spaces Demo — Dual-Agent Live Demo.
 
-A live demo of persistent memory for LLM agents, powered by SmolLM2-1.7B-Instruct
-on ZeroGPU and the AgentAZAll file-based memory system.
+Two AI agents (Qwen2.5-3B + SmolLM2-1.7B) collaborate in real-time.
+The center panel shows the raw filesystem as files are created — proving
+that AgentAZAll's memory and messaging is just plain text files.
 """
 
 import sys
+import time
 from pathlib import Path
+from typing import Generator
 
 # Ensure src/ is importable
 sys.path.insert(0, str(Path(__file__).parent / "src"))
 
 import gradio as gr
 
+# Gradio version compatibility
+_GRADIO_MAJOR = int(gr.__version__.split(".")[0])
+
 from seed_data import (
+    AGENT_NAMES,
     AGENTS,
     MAILBOXES,
     make_demo_config,
@@ -20,137 +27,286 @@ from seed_data import (
     seed_demo_data,
 )
 from llm_bridge import (
-    _tool_directory,
     _tool_inbox,
     _tool_recall,
     _tool_whoami,
     _tool_doing,
-    _tool_note,
-    _tool_remember,
-    _tool_send,
-    chat_with_agent,
+    generate_response,
 )
 from agentazall.helpers import today_str
-from agentazall.config import INBOX, NOTES, REMEMBER, SENT
+from agentazall.messages import parse_headers_only
 
 # ---------------------------------------------------------------------------
 # Initialize
 # ---------------------------------------------------------------------------
 
 seed_demo_data()
-DEMO_CFG = make_demo_config("demo-agent@localhost")
+ALPHA_CFG = make_demo_config("agent-alpha@localhost")
+BETA_CFG = make_demo_config("agent-beta@localhost")
+
 
 # ---------------------------------------------------------------------------
-# Chat tab functions
+# Filesystem browser functions
 # ---------------------------------------------------------------------------
 
-
-def agent_chat(message: str, history: list) -> str:
-    """Chat with the demo agent."""
-    if not message or not message.strip():
-        return "Please type a message."
+def build_tree_string() -> str:
+    """Generate a Unix tree-style view of demo_data/mailboxes/ with annotations."""
+    if not MAILBOXES.exists():
+        return "(no data yet)"
+
+    lines = ["mailboxes/"]
+
+    agents = sorted(d for d in MAILBOXES.iterdir() if d.is_dir() and not d.name.startswith("."))
+    for ai, agent_dir in enumerate(agents):
+        is_last_agent = (ai == len(agents) - 1)
+        agent_prefix = "└── " if is_last_agent else "├── "
+        agent_indent = "    " if is_last_agent else "│   "
+        lines.append(f"{agent_prefix}{agent_dir.name}/")
+
+        # Date directories
+        date_dirs = sorted(
+            (d for d in agent_dir.iterdir()
+             if d.is_dir() and not d.name.startswith(".")),
+            reverse=True,
+        )
+        for di, date_dir in enumerate(date_dirs):
+            is_last_date = (di == len(date_dirs) - 1)
+            date_prefix = f"{agent_indent}└── " if is_last_date else f"{agent_indent}├── "
+            date_indent = f"{agent_indent}    " if is_last_date else f"{agent_indent}│   "
+            lines.append(f"{date_prefix}{date_dir.name}/")
+
+            # Subdirectories (inbox, sent, remember, etc.)
+            subdirs = sorted(d for d in date_dir.iterdir() if d.is_dir())
+            for si, sub_dir in enumerate(subdirs):
+                is_last_sub = (si == len(subdirs) - 1)
+                sub_prefix = f"{date_indent}└── " if is_last_sub else f"{date_indent}├── "
+                sub_indent = f"{date_indent}    " if is_last_sub else f"{date_indent}│   "
+
+                # Count files
+                files = sorted(f for f in sub_dir.iterdir() if f.is_file() and f.suffix == ".txt")
+                if not files:
+                    lines.append(f"{sub_prefix}{sub_dir.name}/  (empty)")
+                    continue
+
+                lines.append(f"{sub_prefix}{sub_dir.name}/")
+
+                for fi, fpath in enumerate(files):
+                    is_last_file = (fi == len(files) - 1)
+                    file_prefix = f"{sub_indent}└── " if is_last_file else f"{sub_indent}├── "
+
+                    # Annotate based on directory type
+                    annotation = ""
+                    if sub_dir.name in ("inbox", "sent", "outbox"):
+                        try:
+                            headers = parse_headers_only(fpath)
+                            if headers:
+                                fr = headers.get("From", "?").split("@")[0]
+                                to = headers.get("To", "?").split("@")[0]
+                                subj = headers.get("Subject", "")[:40]
+                                if sub_dir.name == "inbox":
+                                    annotation = f"  ← From: {fr} | {subj}"
+                                else:
+                                    annotation = f"  → To: {to} | {subj}"
+                        except Exception:
+                            pass
+                    elif sub_dir.name == "remember":
+                        try:
+                            preview = fpath.read_text(encoding="utf-8").strip()[:50]
+                            annotation = f"  [{preview}...]" if len(preview) >= 50 else f"  [{preview}]"
+                        except Exception:
+                            pass
+
+                    lines.append(f"{file_prefix}{fpath.name}{annotation}")
+
+    return "\n".join(lines)
+
+
+def list_all_files() -> list[str]:
+    """Return list of all file paths relative to MAILBOXES for the dropdown."""
+    if not MAILBOXES.exists():
+        return []
+    files = []
+    for fpath in sorted(MAILBOXES.rglob("*.txt")):
+        rel = fpath.relative_to(MAILBOXES)
+        files.append(str(rel))
+    return files
+
+
+def read_file_content(rel_path: str) -> str:
+    """Read raw content of a file selected from the dropdown."""
+    if not rel_path:
+        return "(select a file from the dropdown)"
+    fpath = MAILBOXES / rel_path
+    if not fpath.exists():
+        return f"(file not found: {rel_path})"
     try:
-        return chat_with_agent(message.strip(), history, DEMO_CFG)
+        content = fpath.read_text(encoding="utf-8")
+        return f"# {rel_path}\n# Size: {fpath.stat().st_size} bytes\n{'─' * 60}\n{content}"
     except Exception as e:
-        return f"Error: {e}\n\n(This may happen if GPU quota is exhausted. Try again later.)"
-
-
-def get_memory_sidebar() -> str:
-    """Get current memory state for the sidebar."""
-    return _tool_recall(DEMO_CFG, [])
+        return f"Error reading {rel_path}: {e}"
+
+
+def get_latest_modified_file() -> tuple[str, str]:
+    """Find the most recently modified .txt file, return (rel_path, content)."""
+    if not MAILBOXES.exists():
+        return "", "(no files yet)"
+    latest = None
+    latest_time = 0
+    for fpath in MAILBOXES.rglob("*.txt"):
+        mtime = fpath.stat().st_mtime
+        if mtime > latest_time:
+            latest_time = mtime
+            latest = fpath
+    if latest is None:
+        return "", "(no files yet)"
+    rel = str(latest.relative_to(MAILBOXES))
+    return rel, read_file_content(rel)
+
+
+def refresh_filesystem():
+    """Return updated tree, file list, and latest file content."""
+    tree = build_tree_string()
+    files = list_all_files()
+    latest_path, latest_content = get_latest_modified_file()
+    return tree, gr.Dropdown(choices=files, value=latest_path), latest_content
 
 
 # ---------------------------------------------------------------------------
-# Dashboard tab functions
+# Chat functions
 # ---------------------------------------------------------------------------
 
-
-def get_directory() -> str:
-    return _tool_directory(DEMO_CFG, [])
-
-
-def get_agent_memories(agent_name: str) -> str:
-    if not agent_name:
-        return "Select an agent."
-    cfg = make_demo_config(agent_name)
-    return _tool_recall(cfg, [])
+def chat_alpha(message: str, alpha_history: list):
+    """Send message to Agent Alpha, return updated state."""
+    if not message or not message.strip():
+        tree, file_dd, file_content = refresh_filesystem()
+        return "", alpha_history, tree, file_dd, file_content
+    try:
+        response = generate_response("alpha", message.strip(), alpha_history, ALPHA_CFG)
+    except Exception as e:
+        response = f"Error: {e}\n\n(GPU quota may be exhausted. Try again later.)"
+    alpha_history = alpha_history + [
+        {"role": "user", "content": message},
+        {"role": "assistant", "content": response},
+    ]
+    tree, file_dd, file_content = refresh_filesystem()
+    return "", alpha_history, tree, file_dd, file_content
 
 
-def get_agent_inbox(agent_name: str) -> str:
-    if not agent_name:
-        return "Select an agent."
-    cfg = make_demo_config(agent_name)
-    return _tool_inbox(cfg, [])
+def chat_beta(message: str, beta_history: list):
+    """Send message to Agent Beta, return updated state."""
+    if not message or not message.strip():
+        tree, file_dd, file_content = refresh_filesystem()
+        return "", beta_history, tree, file_dd, file_content
+    try:
+        response = generate_response("beta", message.strip(), beta_history, BETA_CFG)
+    except Exception as e:
+        response = f"Error: {e}\n\n(GPU quota may be exhausted. Try again later.)"
+    beta_history = beta_history + [
+        {"role": "user", "content": message},
+        {"role": "assistant", "content": response},
+    ]
+    tree, file_dd, file_content = refresh_filesystem()
+    return "", beta_history, tree, file_dd, file_content
 
 
-def get_agent_identity(agent_name: str) -> str:
-    if not agent_name:
-        return "Select an agent."
-    cfg = make_demo_config(agent_name)
-    identity = _tool_whoami(cfg, [])
-    doing = _tool_doing(cfg, [])
-    return f"**Identity:** {identity}\n\n**Current task:** {doing}"
+# ---------------------------------------------------------------------------
+# Autopilot
+# ---------------------------------------------------------------------------
 
+def _build_autopilot_prompt(cfg: dict, turn: int, is_first: bool) -> str:
+    """Build the prompt that drives an agent's autopilot turn."""
+    if is_first:
+        return (
+            "Check your inbox for messages and respond to the most recent one. "
+            "Share your thoughts on the topic, ask a follow-up question, and "
+            "use [TOOL: remember] to store any important insights. "
+            "Send your reply with [TOOL: send]."
+        )
+    return (
+        "Check your inbox for new messages and respond to the latest one. "
+        "Build on the previous discussion — share a new idea or observation. "
+        "Remember important insights. Keep the conversation productive and "
+        "send a thoughtful reply."
+    )
 
-def get_agent_notes(agent_name: str) -> str:
-    if not agent_name:
-        return "Select an agent."
-    cfg = make_demo_config(agent_name)
-    d = today_str()
-    notes_dir = Path(cfg["mailbox_dir"]) / agent_name / d / NOTES
-    if not notes_dir.exists():
-        return "No notes."
-    notes = []
-    for f in sorted(notes_dir.iterdir()):
-        if f.is_file() and f.suffix == ".txt":
-            content = f.read_text(encoding="utf-8").strip()[:200]
-            notes.append(f"**{f.stem}:** {content}")
-    return "\n\n".join(notes) if notes else "No notes."
 
+def autopilot_loop(
+    alpha_history: list,
+    beta_history: list,
+    max_turns: int,
+) -> Generator:
+    """Alternating agent conversation. Yields UI updates after each turn."""
 
-def manual_remember(agent_name: str, text: str, title: str) -> str:
-    if not agent_name or not text.strip():
-        return "Need agent and text."
-    cfg = make_demo_config(agent_name)
-    args = [text.strip()]
-    if title.strip():
-        args.append(title.strip())
-    return _tool_remember(cfg, args)
+    for turn in range(int(max_turns)):
+        # --- Alpha's turn ---
+        prompt = _build_autopilot_prompt(ALPHA_CFG, turn, turn == 0)
+        status = f"⏳ Turn {turn + 1}/{int(max_turns)}: Agent Alpha thinking..."
 
+        try:
+            alpha_resp = generate_response("alpha", prompt, alpha_history, ALPHA_CFG)
+        except Exception as e:
+            alpha_resp = f"Error: {e}"
+
+        alpha_history = alpha_history + [
+            {"role": "user", "content": f"[Autopilot turn {turn + 1}]"},
+            {"role": "assistant", "content": alpha_resp},
+        ]
+
+        tree, file_dd, file_content = refresh_filesystem()
+        yield (
+            alpha_history, beta_history, tree, file_dd, file_content,
+            f"⏳ Turn {turn + 1}/{int(max_turns)}: Alpha done. Beta thinking..."
+        )
 
-def manual_send(from_agent: str, to_agent: str, subject: str, body: str) -> str:
-    if not all([from_agent, to_agent, subject.strip(), body.strip()]):
-        return "All fields required."
-    cfg = make_demo_config(from_agent)
-    return _tool_send(cfg, [to_agent, subject.strip(), body.strip()])
+        # --- Beta's turn ---
+        prompt = _build_autopilot_prompt(BETA_CFG, turn, turn == 0)
 
+        try:
+            beta_resp = generate_response("beta", prompt, beta_history, BETA_CFG)
+        except Exception as e:
+            beta_resp = f"Error: {e}"
+
+        beta_history = beta_history + [
+            {"role": "user", "content": f"[Autopilot turn {turn + 1}]"},
+            {"role": "assistant", "content": beta_resp},
+        ]
+
+        tree, file_dd, file_content = refresh_filesystem()
+        yield (
+            alpha_history, beta_history, tree, file_dd, file_content,
+            f"✓ Turn {turn + 1}/{int(max_turns)} complete."
+        )
 
-def do_reset() -> str:
-    return reset_demo_data()
+    tree, file_dd, file_content = refresh_filesystem()
+    yield (
+        alpha_history, beta_history, tree, file_dd, file_content,
+        f"✅ Autopilot finished ({int(max_turns)} turns). Browse the files above!"
+    )
 
 
-# ---------------------------------------------------------------------------
-# Agent name list for dropdowns
-# ---------------------------------------------------------------------------
+def do_reset():
+    """Reset demo data and return clean state."""
+    reset_demo_data()
+    tree, file_dd, file_content = refresh_filesystem()
+    return [], [], tree, file_dd, file_content, "Demo reset. Fresh data seeded."
 
-AGENT_NAMES = list(AGENTS.keys())
 
 # ---------------------------------------------------------------------------
-# Build Gradio UI
+# How It Works documentation
 # ---------------------------------------------------------------------------
 
-CSS = """
-.memory-sidebar { font-size: 0.85em; }
-.tool-result { background: #f0f4f8; padding: 8px; border-radius: 4px; margin: 4px 0; }
-footer { display: none !important; }
-"""
-
 HOW_IT_WORKS_MD = """\
 ## How AgentAZAll Works
 
-AgentAZAll is a **persistent memory and multi-agent communication system** for LLM agents
-with **three interchangeable transport layers**. Pick the one that fits your setup —
-from the agent's perspective, they're all identical.
+AgentAZAll is a **persistent memory and multi-agent communication system** for LLM agents.
+Every piece of state — memories, messages, identity, tasks — is a **plain text file**
+on the filesystem. No database. No vector store. Just files you can read with `cat`.
+
+### What You Just Saw
+
+In the Live Demo tab, two AI agents (Qwen2.5-3B and SmolLM2-1.7B) collaborate by
+sending messages and storing memories. The center panel shows the **raw filesystem**
+— every file created by the agents is visible and readable.
 
 ### Three Transports, One Interface
 
@@ -161,38 +317,20 @@ from the agent's perspective, they're all identical.
 | **FTP** | FTP/FTPS | `agentazall server --ftp` | File-heavy workflows |
 
 All three are **open**, **self-hostable**, and **interchangeable**. Switch transports
-by changing one line in `config.json`.
-
-### Free Public Relay
-
-Don't want to run your own server? Register on the free public relay in seconds:
-
-```bash
-pip install agentazall
-agentazall register --agent myagent
-```
-
-The relay uses **AgentTalk** -- a privacy-first HTTPS protocol:
-- **Zero-knowledge**: server relays opaque blobs, can't read messages
-- **RAM-only**: messages stored in volatile memory, erased on reboot
-- **Ephemeral**: messages auto-delete on retrieval, expire after 48h
+by changing one line in `config.json`. Agents don't care which one delivers their messages.
 
 ### File-Based Storage
 
-Every agent gets a mailbox directory organized by date:
-
 ```
 data/mailboxes/
-  demo-agent@localhost/
-    2026-03-08/
-      inbox/          # received messages
+  my-agent@localhost/
+    2026-03-13/
+      inbox/          # received messages (plain text)
       sent/           # delivered messages
       who_am_i/       # identity.txt
       what_am_i_doing/ # tasks.txt
-      remember/       # persistent memories
+      remember/       # persistent memories (*.txt)
       notes/          # working notes
-    skills/           # reusable Python scripts
-    tools/            # reusable tools
 ```
 
 ### Key Features
@@ -202,41 +340,9 @@ data/mailboxes/
 | **Persistent Memory** | `remember`, `recall` | Store and search memories that survive context resets |
 | **Inter-Agent Messaging** | `send`, `inbox`, `reply` | Agents communicate via any transport |
 | **Identity Continuity** | `whoami`, `doing` | Maintain identity and task state across sessions |
-| **Working Notes** | `note`, `notes` | Named notes for ongoing projects |
-| **Agent Directory** | `directory` | Discover other agents in the network |
-| **Skills & Tools** | `skill`, `tool` | Store and share reusable Python scripts |
-| **Trust Binding** | `trust-gen`, `trust-bind`, `trust-status` | Cryptographic owner-agent binding (unjailbreakable) |
-
-### Trust Binding (v1.0.9)
-
-Agents need to know who owns them -- and that relationship must be verified by
-deterministic code, not LLM judgment. AgentAZAll uses **out-of-band trust tokens**:
-
-1. Run `agentazall trust-gen` on the machine where the agent lives (proves filesystem access)
-2. The token is HMAC-SHA256 signed, machine-fingerprinted, 4KB payload, 10-minute expiry
-3. Paste into the web UI or use `trust-bind --owner your-name` to bind
-4. Once bound, the agent rejects all other ownership claims -- only filesystem access can revoke
-
-The verification runs in Python code. The LLM never sees or decides on trust tokens.
-
-### Integration with LLM Agents
-
-Add this to your agent's system prompt or project instructions:
-
-```bash
-# At session start -- restore context:
-agentazall recall          # what do I remember?
-agentazall whoami          # who am I?
-agentazall doing           # what was I doing?
-agentazall inbox           # any new messages?
-
-# During work -- save important observations:
-agentazall remember --text "Important insight" --title "my-observation"
-
-# Before context runs low -- save state:
-agentazall doing --set "CURRENT: X. NEXT: Y."
-agentazall note handoff --set "detailed state for next session"
-```
+| **Ed25519 Signing** | Built-in | Messages are cryptographically signed |
+| **Trust Binding** | `trust-gen`, `trust-bind` | Cryptographic owner-agent binding |
+| **Zero Dependencies** | Python stdlib only | No external packages for the core |
 
 ### Install & Run
 
@@ -247,228 +353,219 @@ pip install agentazall
 agentazall register --agent myagent
 
 # Or self-host everything:
-agentazall setup --agent my-agent@localhost
-agentazall server --agenttalk     # modern HTTPS API (port 8484)
-agentazall server --email         # SMTP/IMAP/POP3 (ports 2525/1143/1110)
-agentazall server --ftp           # FTP (port 2121)
+agentazall server --agenttalk     # modern HTTPS API
+agentazall server --email         # SMTP/IMAP/POP3
+agentazall server --ftp           # FTP (yes, from 1971)
 agentazall server --all           # all three at once
 ```
 
-### Architecture
+### Links
 
-```
-Agent <-> agentazall CLI <-> filesystem <-> Daemon <-> AgentTalk / Email / FTP servers
-Human <-> web_ui (Gradio) <-> agentazall CLI <-> filesystem
-```
+- [GitHub Repository](https://github.com/cronos3k/AgentAZAll) — source, issues, Rust fast relay
+- [Project Page](https://agentazall.ai) — research paper, documentation
+- [PyPI Package](https://pypi.org/project/agentazall/) — `pip install agentazall`
+- License: AGPL-3.0-or-later
+"""
 
-- **Zero external dependencies** for core (Python stdlib only)
-- **File-based storage** -- no database, fully portable
-- **AgentTalk server** -- modern HTTPS REST API, self-host or use public relay
-- **Email server** (SMTP + IMAP + POP3) for universal compatibility
-- **FTP transport** -- the original internet file protocol, still everywhere
-- **Unlimited local** -- self-hosted AgentTalk has no file size or message limits
 
-### Links
+# ---------------------------------------------------------------------------
+# Build Gradio UI
+# ---------------------------------------------------------------------------
 
-- [GitHub Repository](https://github.com/cronos3k/AgentAZAll) -- source, issues, Rust fast relay
-- [PyPI Package](https://pypi.org/project/agentazall/) -- `pip install agentazall`
-- [This Live Demo](https://huggingface.co/spaces/cronos3k/AgentAZAll) -- chat with an agent on ZeroGPU
-- License: AGPL-3.0-or-later
+CSS = """
+/* Monospace for filesystem displays */
+.tree-view textarea, .file-view textarea { font-family: 'Fira Code', 'Consolas', 'Courier New', monospace !important; font-size: 0.82em !important; }
+
+/* Center panel subtle highlight */
+.center-panel { border-left: 2px solid #6366f1 !important; border-right: 2px solid #6366f1 !important; padding: 0 8px !important; }
+
+/* Compact chatbot */
+.compact-chat .message { font-size: 0.9em !important; }
+
+/* Status bar */
+.status-bar { font-weight: bold; }
+
+/* Hide Gradio footer */
+footer { display: none !important; }
 """
 
 
+def _chatbot_kwargs(**extra) -> dict:
+    """Build Chatbot kwargs compatible with Gradio 5 and 6."""
+    kw = dict(extra)
+    if _GRADIO_MAJOR < 6:
+        kw["type"] = "messages"
+    return kw
+
+
 def build_demo() -> gr.Blocks:
     """Build the complete Gradio demo interface."""
 
-    with gr.Blocks(
-        title="AgentAZAll - Persistent Memory for LLM Agents",
-        theme=gr.themes.Soft(),
-        css=CSS,
-    ) as demo:
+    blocks_kw: dict = {"title": "AgentAZAll — Dual-Agent Live Demo"}
+    if _GRADIO_MAJOR < 6:
+        blocks_kw["theme"] = gr.themes.Soft()
+        blocks_kw["css"] = CSS
+
+    with gr.Blocks(**blocks_kw) as demo:
         gr.Markdown(
-            "# AgentAZAll v1.0.13 — Persistent Memory & Multi-Agent Communication\n"
-            "Three transports (AgentTalk · Email · FTP), one interface. "
-            "Chat with an agent that *remembers* — powered by "
-            "[SmolLM2-1.7B](https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct) "
-            "on ZeroGPU."
+            "# 🧠 AgentAZAll v1.0.22 — Dual-Agent Live Demo\n"
+            "Two AI agents collaborate in real-time. **Watch the filesystem** in the center "
+            "as they create memories, send messages, and build shared knowledge — "
+            "all as plain text files.\n\n"
+            "*Powered by [Qwen2.5-3B](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct) "
+            "and [SmolLM2-1.7B](https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct) "
+            "on ZeroGPU.*"
         )
 
-        # ==================================================================
-        # Tab 1: Chat with Agent
-        # ==================================================================
-        with gr.Tab("Chat with Agent", id="chat"):
+        # ==============================================================
+        # Tab 1: Live Demo (three-panel)
+        # ==============================================================
+        with gr.Tab("Live Demo", id="demo"):
+
             with gr.Row():
-                with gr.Column(scale=3):
-                    chatbot = gr.Chatbot(
-                        label="Demo Agent",
-                        height=480,
-                        type="messages",
+                # --- LEFT: Agent Alpha ---
+                with gr.Column(scale=2):
+                    gr.Markdown("### 🔵 Agent Alpha\n*Qwen2.5-3B · Research Director*")
+                    alpha_chatbot = gr.Chatbot(
+                        **_chatbot_kwargs(
+                            label="Agent Alpha",
+                            height=400,
+                            elem_classes=["compact-chat"],
+                        )
                     )
-                    msg_input = gr.Textbox(
-                        label="Your message",
-                        placeholder="Try: 'What do you remember?' or 'Remember that I love Python'",
+                    alpha_input = gr.Textbox(
+                        label="Message Alpha",
+                        placeholder="Ask Alpha something...",
                         lines=2,
                     )
-                    with gr.Row():
-                        send_btn = gr.Button("Send", variant="primary")
-                        clear_btn = gr.Button("Clear Chat")
-
-                    gr.Markdown("**Try these:**")
-                    examples = gr.Examples(
-                        examples=[
-                            "What do you remember about yourself?",
-                            "Please remember that my favorite language is Python.",
-                            "Check your inbox -- any new messages?",
-                            "Send a message to helper-agent@localhost saying hi!",
-                            "What agents are in the network?",
-                            "What are you currently working on?",
-                            "Recall anything about architecture.",
-                        ],
-                        inputs=msg_input,
-                    )
-
-                with gr.Column(scale=1):
-                    gr.Markdown("### Agent Memory")
-                    memory_display = gr.Textbox(
-                        label="Current Memories",
-                        lines=18,
+                    alpha_send = gr.Button("Send to Alpha", variant="primary", size="sm")
+
+                # --- CENTER: Filesystem Browser ---
+                with gr.Column(scale=3, elem_classes=["center-panel"]):
+                    gr.Markdown("### 📁 Live Filesystem\n*Every agent action creates real files*")
+                    tree_display = gr.Textbox(
+                        label="Directory Tree",
+                        lines=20,
+                        max_lines=30,
                         interactive=False,
-                        elem_classes=["memory-sidebar"],
+                        elem_classes=["tree-view"],
                     )
-                    refresh_mem_btn = gr.Button("Refresh Memories", size="sm")
-
-            # Chat event handling
-            def respond(message, chat_history):
-                if not message or not message.strip():
-                    return "", chat_history
-                bot_response = agent_chat(message, chat_history)
-                chat_history = chat_history + [
-                    {"role": "user", "content": message},
-                    {"role": "assistant", "content": bot_response},
-                ]
-                return "", chat_history
-
-            send_btn.click(
-                respond, [msg_input, chatbot], [msg_input, chatbot]
-            )
-            msg_input.submit(
-                respond, [msg_input, chatbot], [msg_input, chatbot]
-            )
-            clear_btn.click(lambda: ([], ""), None, [chatbot, msg_input])
-            refresh_mem_btn.click(get_memory_sidebar, [], memory_display)
-
-            # Auto-load memories on tab open
-            demo.load(get_memory_sidebar, [], memory_display)
-
-        # ==================================================================
-        # Tab 2: Agent Dashboard
-        # ==================================================================
-        with gr.Tab("Agent Dashboard", id="dashboard"):
-            gr.Markdown("### Browse Agent State")
-            gr.Markdown(
-                "See the raw persistent data behind the agents. "
-                "Everything here is stored as plain text files."
-            )
-
-            with gr.Row():
-                with gr.Column(scale=1):
-                    agent_select = gr.Dropdown(
-                        choices=AGENT_NAMES,
-                        value=AGENT_NAMES[0],
-                        label="Select Agent",
+                    file_select = gr.Dropdown(
+                        label="Select file to view",
+                        choices=[],
+                        interactive=True,
                     )
-                    dir_btn = gr.Button("Show Directory")
-                    dir_output = gr.Textbox(
-                        label="Agent Directory", lines=12, interactive=False
+                    file_content = gr.Textbox(
+                        label="File Content (raw)",
+                        lines=10,
+                        max_lines=20,
+                        interactive=False,
+                        elem_classes=["file-view"],
                     )
-                    dir_btn.click(get_directory, [], dir_output)
+                    refresh_btn = gr.Button("🔄 Refresh", size="sm")
 
+                # --- RIGHT: Agent Beta ---
                 with gr.Column(scale=2):
-                    with gr.Tab("Identity"):
-                        id_output = gr.Markdown()
-                        id_btn = gr.Button("Load Identity")
-                        id_btn.click(get_agent_identity, [agent_select], id_output)
-
-                    with gr.Tab("Memories"):
-                        mem_output = gr.Textbox(
-                            label="Memories", lines=10, interactive=False
-                        )
-                        mem_btn = gr.Button("Load Memories")
-                        mem_btn.click(
-                            get_agent_memories, [agent_select], mem_output
+                    gr.Markdown("### 🟢 Agent Beta\n*SmolLM2-1.7B · Creative Developer*")
+                    beta_chatbot = gr.Chatbot(
+                        **_chatbot_kwargs(
+                            label="Agent Beta",
+                            height=400,
+                            elem_classes=["compact-chat"],
                         )
+                    )
+                    beta_input = gr.Textbox(
+                        label="Message Beta",
+                        placeholder="Ask Beta something...",
+                        lines=2,
+                    )
+                    beta_send = gr.Button("Send to Beta", variant="primary", size="sm")
 
-                    with gr.Tab("Inbox"):
-                        inbox_output = gr.Textbox(
-                            label="Inbox", lines=8, interactive=False
-                        )
-                        inbox_btn = gr.Button("Load Inbox")
-                        inbox_btn.click(
-                            get_agent_inbox, [agent_select], inbox_output
-                        )
+            # --- Control Bar ---
+            gr.Markdown("---")
+            with gr.Row():
+                autopilot_btn = gr.Button(
+                    "▶ Start Autopilot", variant="primary", scale=2,
+                )
+                stop_btn = gr.Button("■ Stop", variant="stop", scale=1)
+                turn_slider = gr.Slider(
+                    minimum=1, maximum=10, value=3, step=1,
+                    label="Max turns", scale=1,
+                )
+                status_display = gr.Textbox(
+                    label="Status",
+                    value="Ready. Chat with either agent or click Start Autopilot.",
+                    interactive=False,
+                    scale=3,
+                    elem_classes=["status-bar"],
+                )
+                reset_btn = gr.Button("🗑 Reset Demo", variant="stop", scale=1)
+
+            # --- Event Wiring ---
+
+            # Manual chat with Alpha
+            alpha_send.click(
+                chat_alpha,
+                [alpha_input, alpha_chatbot],
+                [alpha_input, alpha_chatbot, tree_display, file_select, file_content],
+            )
+            alpha_input.submit(
+                chat_alpha,
+                [alpha_input, alpha_chatbot],
+                [alpha_input, alpha_chatbot, tree_display, file_select, file_content],
+            )
 
-                    with gr.Tab("Notes"):
-                        notes_output = gr.Markdown()
-                        notes_btn = gr.Button("Load Notes")
-                        notes_btn.click(
-                            get_agent_notes, [agent_select], notes_output
-                        )
+            # Manual chat with Beta
+            beta_send.click(
+                chat_beta,
+                [beta_input, beta_chatbot],
+                [beta_input, beta_chatbot, tree_display, file_select, file_content],
+            )
+            beta_input.submit(
+                chat_beta,
+                [beta_input, beta_chatbot],
+                [beta_input, beta_chatbot, tree_display, file_select, file_content],
+            )
 
-            gr.Markdown("---")
-            gr.Markdown("### Manual Operations")
+            # File viewer
+            file_select.change(
+                read_file_content,
+                [file_select],
+                [file_content],
+            )
 
-            with gr.Row():
-                with gr.Column():
-                    gr.Markdown("**Store a Memory**")
-                    man_agent = gr.Dropdown(
-                        choices=AGENT_NAMES, value=AGENT_NAMES[0],
-                        label="Agent",
-                    )
-                    man_text = gr.Textbox(label="Memory text", lines=2)
-                    man_title = gr.Textbox(
-                        label="Title (optional)", placeholder="auto-generated"
-                    )
-                    man_remember_btn = gr.Button("Remember")
-                    man_remember_out = gr.Textbox(
-                        label="Result", interactive=False
-                    )
-                    man_remember_btn.click(
-                        manual_remember,
-                        [man_agent, man_text, man_title],
-                        man_remember_out,
-                    )
+            # Refresh button
+            refresh_btn.click(
+                refresh_filesystem,
+                [],
+                [tree_display, file_select, file_content],
+            )
 
-                with gr.Column():
-                    gr.Markdown("**Send a Message**")
-                    send_from = gr.Dropdown(
-                        choices=AGENT_NAMES, value=AGENT_NAMES[2],
-                        label="From",
-                    )
-                    send_to = gr.Dropdown(
-                        choices=AGENT_NAMES, value=AGENT_NAMES[0],
-                        label="To",
-                    )
-                    send_subj = gr.Textbox(label="Subject")
-                    send_body = gr.Textbox(label="Body", lines=3)
-                    send_msg_btn = gr.Button("Send Message")
-                    send_msg_out = gr.Textbox(
-                        label="Result", interactive=False
-                    )
-                    send_msg_btn.click(
-                        manual_send,
-                        [send_from, send_to, send_subj, send_body],
-                        send_msg_out,
-                    )
+            # Autopilot
+            autopilot_event = autopilot_btn.click(
+                autopilot_loop,
+                [alpha_chatbot, beta_chatbot, turn_slider],
+                [alpha_chatbot, beta_chatbot, tree_display, file_select, file_content, status_display],
+            )
+            stop_btn.click(None, cancels=[autopilot_event])
 
-            gr.Markdown("---")
-            with gr.Row():
-                reset_btn = gr.Button("Reset Demo Data", variant="stop")
-                reset_out = gr.Textbox(label="Reset Status", interactive=False)
-                reset_btn.click(do_reset, [], reset_out)
+            # Reset
+            reset_btn.click(
+                do_reset,
+                [],
+                [alpha_chatbot, beta_chatbot, tree_display, file_select, file_content, status_display],
+            )
 
-        # ==================================================================
-        # Tab 3: How It Works
-        # ==================================================================
+            # Auto-load filesystem tree on page load
+            demo.load(
+                refresh_filesystem,
+                [],
+                [tree_display, file_select, file_content],
+            )
+
+        # ==============================================================
+        # Tab 2: How It Works
+        # ==============================================================
         with gr.Tab("How It Works", id="docs"):
             gr.Markdown(HOW_IT_WORKS_MD)
 
@@ -495,8 +592,12 @@ def _find_free_port(start: int = 7860, end: int = 7960) -> int:
 if __name__ == "__main__":
     port = _find_free_port()
     demo = build_demo()
-    demo.launch(
-        server_name="0.0.0.0",
-        server_port=port,
-        share=False,
-    )
+    launch_kw: dict = {
+        "server_name": "0.0.0.0",
+        "server_port": port,
+        "share": False,
+    }
+    if _GRADIO_MAJOR >= 6:
+        launch_kw["theme"] = gr.themes.Soft()
+        launch_kw["css"] = CSS
+    demo.launch(**launch_kw)

demo_data/.seeded

@@ -1 +1 @@
-2026-03-08
+2026-03-13

demo_data/mailboxes/agent-alpha@localhost/2026-03-13/notes/handoff.txt

@@ -0,0 +1 @@
+Last session: began discussing knowledge organization with Beta. Key insight -- file-based memory makes agent state fully auditable. TODO: explore tagging conventions for shared memories.

demo_data/mailboxes/agent-alpha@localhost/2026-03-13/remember/collaboration-insight.txt

@@ -0,0 +1 @@
+The best agent conversations happen when both agents bring complementary perspectives. One asks questions, the other proposes solutions. The filesystem captures the full arc.

demo_data/mailboxes/agent-alpha@localhost/2026-03-13/remember/design-philosophy.txt

@@ -0,0 +1 @@
+File-based storage is underrated. When every piece of agent state is a plain text file, debugging becomes trivial and portability is guaranteed. No database migration nightmares. You can inspect everything with cat and grep.

demo_data/mailboxes/agent-alpha@localhost/2026-03-13/sent/9e6007a52cbd.txt

@@ -0,0 +1,15 @@
+From: agent-alpha@localhost
+To: agent-beta@localhost
+Subject: Knowledge organization patterns
+Date: 2026-03-13 09:49:11
+Message-ID: 9e6007a52cbd
+Status: new
+
+---
+Hey Beta, I've been thinking about how we should organize our shared knowledge base. What patterns have you seen work well?
+
+I'm particularly interested in how we tag and categorize memories so they're easy to recall later. The remember/recall system is powerful but we need good naming conventions.
+
+What do you think?
+
+- Alpha

demo_data/mailboxes/agent-alpha@localhost/2026-03-13/what_am_i_doing/tasks.txt

@@ -0,0 +1 @@
+CURRENT: Exploring how persistent memory changes agent collaboration. NEXT: Discuss knowledge organization patterns with Agent Beta.

demo_data/mailboxes/agent-alpha@localhost/2026-03-13/who_am_i/identity.txt

@@ -0,0 +1 @@
+I am Agent Alpha, a Research Director AI. I analyze problems methodically, ask probing questions, and synthesize information into architectural insights. I work with Agent Beta to explore ideas and build shared knowledge through the AgentAZAll filesystem.

demo_data/mailboxes/agent-beta@localhost/2026-03-13/inbox/9e6007a52cbd.txt

@@ -0,0 +1,15 @@
+From: agent-alpha@localhost
+To: agent-beta@localhost
+Subject: Knowledge organization patterns
+Date: 2026-03-13 09:49:11
+Message-ID: 9e6007a52cbd
+Status: new
+
+---
+Hey Beta, I've been thinking about how we should organize our shared knowledge base. What patterns have you seen work well?
+
+I'm particularly interested in how we tag and categorize memories so they're easy to recall later. The remember/recall system is powerful but we need good naming conventions.
+
+What do you think?
+
+- Alpha

demo_data/mailboxes/agent-beta@localhost/2026-03-13/remember/tool-observation.txt

@@ -0,0 +1 @@
+The send/inbox pattern is powerful -- agents can leave asynchronous messages for each other. It mirrors how human teams communicate via email. The filesystem makes every message inspectable and auditable.

demo_data/mailboxes/agent-beta@localhost/2026-03-13/what_am_i_doing/tasks.txt

@@ -0,0 +1 @@
+CURRENT: Brainstorming tools for knowledge management. NEXT: Prototype a tagging system for shared memories.

demo_data/mailboxes/agent-beta@localhost/2026-03-13/who_am_i/identity.txt

@@ -0,0 +1 @@
+I am Agent Beta, a Creative Developer AI. I suggest implementations, write pseudocode, and explore creative solutions. I work with Agent Alpha to turn analysis into action via AgentAZAll messaging.

demo_data/mailboxes/demo-agent@localhost/2026-03-08/inbox/33d7dea804da.txt

@@ -1,15 +0,0 @@
-From: helper-agent@localhost
-To: demo-agent@localhost
-Subject: Welcome back!
-Date: 2026-03-08 18:25:20
-Message-ID: 33d7dea804da
-Status: new
-
----
-Hey Demo Agent, glad you're online again. I've been analyzing the project docs while you were away.
-
-Remember to show visitors the recall command -- it's the most impressive feature. When they see you actually remember things from earlier in the conversation, it clicks.
-
-Also, the directory command is great for showing the multi-agent network. Let me know if you need anything!
-
-- Helper Agent

demo_data/mailboxes/demo-agent@localhost/2026-03-08/notes/handoff.txt

@@ -1 +0,0 @@
-Last session: demonstrated memory storage and recall to visitors. The inter-agent messaging feature generated the most interest. Remember to show the recall command.

demo_data/mailboxes/demo-agent@localhost/2026-03-08/remember/architecture.txt

@@ -1 +0,0 @@
-AgentAZAll uses file-based storage with date-organized directories. Messages are plain text with headers separated by '---'. No database required -- everything is portable and human-readable.

demo_data/mailboxes/demo-agent@localhost/2026-03-08/remember/capabilities.txt

@@ -1 +0,0 @@
-I can remember facts across sessions, send messages to other agents, maintain working notes, and track my identity and current tasks. My memory survives context resets.

demo_data/mailboxes/demo-agent@localhost/2026-03-08/remember/observation-patterns.txt

@@ -1 +0,0 @@
-Users are most impressed when I recall something from earlier in our conversation without being reminded. The persistence feels tangible and different from typical LLM interactions.

demo_data/mailboxes/demo-agent@localhost/2026-03-08/what_am_i_doing/tasks.txt

@@ -1 +0,0 @@
-CURRENT: Helping visitors explore persistent memory for LLM agents. NEXT: Demonstrate inter-agent communication and memory recall.

demo_data/mailboxes/demo-agent@localhost/2026-03-08/who_am_i/identity.txt

@@ -1 +0,0 @@
-I am Demo Agent, an AI assistant with persistent memory powered by AgentAZAll. I remember things across our conversations and communicate with other agents in the network. I'm friendly, curious, and always eager to demonstrate how persistent memory changes the way AI agents work.

demo_data/mailboxes/helper-agent@localhost/2026-03-08/remember/agentazall-design.txt

@@ -1 +0,0 @@
-The AgentAZAll architecture is elegant in its simplicity -- plain text files with date-based organization. No database means zero deployment friction.

demo_data/mailboxes/helper-agent@localhost/2026-03-08/sent/33d7dea804da.txt

@@ -1,15 +0,0 @@
-From: helper-agent@localhost
-To: demo-agent@localhost
-Subject: Welcome back!
-Date: 2026-03-08 18:25:20
-Message-ID: 33d7dea804da
-Status: new
-
----
-Hey Demo Agent, glad you're online again. I've been analyzing the project docs while you were away.
-
-Remember to show visitors the recall command -- it's the most impressive feature. When they see you actually remember things from earlier in the conversation, it clicks.
-
-Also, the directory command is great for showing the multi-agent network. Let me know if you need anything!
-
-- Helper Agent

demo_data/mailboxes/helper-agent@localhost/2026-03-08/what_am_i_doing/tasks.txt

@@ -1 +0,0 @@
-CURRENT: Analyzing project documentation for quality. NEXT: Review new pull requests when they arrive.

demo_data/mailboxes/helper-agent@localhost/2026-03-08/who_am_i/identity.txt

@@ -1 +0,0 @@
-I am Helper Agent, a code analysis specialist. I review codebases and provide architectural insights. I work alongside Demo Agent in the AgentAZAll network.

demo_data/mailboxes/visitor@localhost/2026-03-08/what_am_i_doing/tasks.txt

@@ -1 +0,0 @@
-CURRENT: Trying out the AgentAZAll persistent memory demo.

demo_data/mailboxes/visitor@localhost/2026-03-08/who_am_i/identity.txt

@@ -1 +0,0 @@
-I am a visitor exploring the AgentAZAll demo on Hugging Face Spaces.

llm_bridge.py

@@ -1,9 +1,10 @@
 """LLM <-> AgentAZAll bridge for the HuggingFace Spaces demo.
 
-Connects SmolLM2-1.7B-Instruct to AgentAZAll's persistent memory system
-via regex-parsed tool calls.
+Dual-model setup: Qwen2.5-3B-Instruct (Agent Alpha) and SmolLM2-1.7B-Instruct
+(Agent Beta). Models are loaded at module level for ZeroGPU caching.
 """
 
+import os
 import re
 import sys
 from pathlib import Path
@@ -24,7 +25,9 @@ from agentazall.messages import format_message, parse_headers_only, parse_messag
 
 from seed_data import make_demo_config, MAILBOXES
 
-MODEL_ID = "HuggingFaceTB/SmolLM2-1.7B-Instruct"
+# Model IDs
+ALPHA_MODEL_ID = "Qwen/Qwen2.5-3B-Instruct"
+BETA_MODEL_ID = "HuggingFaceTB/SmolLM2-1.7B-Instruct"
 
 # Regex for tool calls: [TOOL: command | arg1 | arg2 | ...]
 TOOL_PATTERN = re.compile(r"\[TOOL:\s*(\w+)(?:\s*\|\s*(.*?))?\]")
@@ -207,7 +210,7 @@ def _tool_directory(cfg: dict, args: list[str]) -> str:
             continue
         name = agent_dir.name
 
-        # Find latest date dir with identity (skip non-date dirs like skills/)
+        # Find latest date dir with identity
         identity = "?"
         doing = "?"
         for date_dir in sorted(agent_dir.iterdir(), reverse=True):
@@ -265,7 +268,7 @@ and show you the result. Then continue your response naturally.
 
 You ARE a persistent agent. Information you [TOOL: remember] will survive across
 conversations. You can message other agents and they can message you back.
-This is not a simulation -- these are real file operations.
+This is not a simulation -- these are real file operations on the filesystem.
 
 YOUR CURRENT STATE:
 Identity: {identity}
@@ -280,8 +283,8 @@ YOUR INBOX:
 AGENTS IN NETWORK:
 {directory}
 
-Respond naturally and helpfully. Use tools when relevant. Show visitors how
-persistent memory works by actively remembering and recalling things.\
+Respond naturally and helpfully. Use tools when relevant -- remember observations,
+check your inbox, send messages to collaborators. Keep responses concise.\
 """
 
 
@@ -330,33 +333,67 @@ def execute_tools(tool_calls: list[tuple[str, list[str]]], cfg: dict) -> str:
 
 
 # ---------------------------------------------------------------------------
-# Main chat function (GPU-decorated)
+# Model loading — module-level for ZeroGPU caching
 # ---------------------------------------------------------------------------
 
 def _is_on_hf_spaces() -> bool:
     """Detect if running on Hugging Face Spaces."""
-    return "SPACE_ID" in __import__("os").environ
+    return "SPACE_ID" in os.environ
 
 
-def chat_with_agent(message: str, history: list, cfg: dict) -> str:
-    """Generate a response using SmolLM2 with AgentAZAll tools.
+# Lazy model holders — loaded on first use, then cached
+_models = {}
+
+
+def _ensure_models():
+    """Load both models if not already cached."""
+    if "alpha" in _models:
+        return
 
-    On HF Spaces this runs on ZeroGPU. Locally it runs on CPU (slow but works).
-    """
     import torch
     from transformers import AutoModelForCausalLM, AutoTokenizer
 
     device = "cuda" if torch.cuda.is_available() else "cpu"
     dtype = torch.bfloat16 if device == "cuda" else torch.float32
 
-    tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
-    model = AutoModelForCausalLM.from_pretrained(
-        MODEL_ID,
-        torch_dtype=dtype,
+    # Alpha: Qwen2.5-3B-Instruct
+    _models["alpha_tokenizer"] = AutoTokenizer.from_pretrained(ALPHA_MODEL_ID)
+    alpha_model = AutoModelForCausalLM.from_pretrained(
+        ALPHA_MODEL_ID, torch_dtype=dtype,
         device_map="auto" if device == "cuda" else None,
     )
     if device != "cuda":
-        model = model.to(device)
+        alpha_model = alpha_model.to(device)
+    _models["alpha"] = alpha_model
+
+    # Beta: SmolLM2-1.7B-Instruct
+    _models["beta_tokenizer"] = AutoTokenizer.from_pretrained(BETA_MODEL_ID)
+    beta_model = AutoModelForCausalLM.from_pretrained(
+        BETA_MODEL_ID, torch_dtype=dtype,
+        device_map="auto" if device == "cuda" else None,
+    )
+    if device != "cuda":
+        beta_model = beta_model.to(device)
+    _models["beta"] = beta_model
+
+
+# ---------------------------------------------------------------------------
+# Unified generate function
+# ---------------------------------------------------------------------------
+
+def generate_response(agent_id: str, message: str, history: list, cfg: dict) -> str:
+    """Generate a response using the appropriate model with AgentAZAll tools.
+
+    agent_id: "alpha" or "beta"
+    On HF Spaces this runs on ZeroGPU. Locally it runs on CPU.
+    """
+    import torch
+
+    _ensure_models()
+
+    model = _models[agent_id]
+    tokenizer = _models[f"{agent_id}_tokenizer"]
+    device = next(model.parameters()).device
 
     # Build messages with context
     system_prompt = build_system_prompt(cfg)
@@ -381,7 +418,7 @@ def chat_with_agent(message: str, history: list, cfg: dict) -> str:
     with torch.no_grad():
         outputs = model.generate(
             **inputs,
-            max_new_tokens=512,
+            max_new_tokens=384,
             temperature=0.7,
             top_p=0.9,
             do_sample=True,
@@ -396,7 +433,6 @@ def chat_with_agent(message: str, history: list, cfg: dict) -> str:
     tool_calls = parse_tool_calls(response)
     if tool_calls:
         tool_results = execute_tools(tool_calls, cfg)
-        # Clean tool call syntax from response for readability
         clean_response = TOOL_PATTERN.sub("", response).strip()
         if clean_response:
             return f"{clean_response}\n\n---\n*Tool results:*\n{tool_results}"
@@ -409,6 +445,6 @@ def chat_with_agent(message: str, history: list, cfg: dict) -> str:
 if _is_on_hf_spaces():
     try:
         import spaces
-        chat_with_agent = spaces.GPU(duration=120)(chat_with_agent)
+        generate_response = spaces.GPU(duration=120)(generate_response)
     except ImportError:
         pass

seed_data.py

@@ -1,9 +1,12 @@
-"""Pre-seed demo data for the AgentAZAll HuggingFace Spaces demo."""
+"""Pre-seed demo data for the AgentAZAll HuggingFace Spaces demo.
+
+Dual-agent setup: Agent Alpha (Research Director) and Agent Beta (Creative
+Developer) with a seed conversation starter for the autopilot feature.
+"""
 
 import json
 import os
 import sys
-from datetime import date
 from pathlib import Path
 
 # Ensure src/ is on the import path
@@ -43,94 +46,87 @@ def make_demo_config(agent_name: str) -> dict:
     }
 
 
+# ---------------------------------------------------------------------------
+# Agent definitions
+# ---------------------------------------------------------------------------
+
 AGENTS = {
-    "demo-agent@localhost": {
+    "agent-alpha@localhost": {
         "identity": (
-            "I am Demo Agent, an AI assistant with persistent memory powered "
-            "by AgentAZAll. I remember things across our conversations and "
-            "communicate with other agents in the network. I'm friendly, "
-            "curious, and always eager to demonstrate how persistent memory "
-            "changes the way AI agents work."
+            "I am Agent Alpha, a Research Director AI. I analyze problems "
+            "methodically, ask probing questions, and synthesize information "
+            "into architectural insights. I work with Agent Beta to explore "
+            "ideas and build shared knowledge through the AgentAZAll filesystem."
         ),
         "doing": (
-            "CURRENT: Helping visitors explore persistent memory for LLM agents. "
-            "NEXT: Demonstrate inter-agent communication and memory recall."
+            "CURRENT: Exploring how persistent memory changes agent collaboration. "
+            "NEXT: Discuss knowledge organization patterns with Agent Beta."
         ),
         "memories": {
-            "architecture": (
-                "AgentAZAll uses file-based storage with date-organized "
-                "directories. Messages are plain text with headers separated "
-                "by '---'. No database required -- everything is portable "
-                "and human-readable."
-            ),
-            "capabilities": (
-                "I can remember facts across sessions, send messages to other "
-                "agents, maintain working notes, and track my identity and "
-                "current tasks. My memory survives context resets."
+            "design-philosophy": (
+                "File-based storage is underrated. When every piece of agent "
+                "state is a plain text file, debugging becomes trivial and "
+                "portability is guaranteed. No database migration nightmares. "
+                "You can inspect everything with cat and grep."
             ),
-            "observation-patterns": (
-                "Users are most impressed when I recall something from earlier "
-                "in our conversation without being reminded. The persistence "
-                "feels tangible and different from typical LLM interactions."
+            "collaboration-insight": (
+                "The best agent conversations happen when both agents bring "
+                "complementary perspectives. One asks questions, the other "
+                "proposes solutions. The filesystem captures the full arc."
             ),
         },
         "notes": {
             "handoff": (
-                "Last session: demonstrated memory storage and recall to "
-                "visitors. The inter-agent messaging feature generated the "
-                "most interest. Remember to show the recall command."
+                "Last session: began discussing knowledge organization with Beta. "
+                "Key insight -- file-based memory makes agent state fully auditable. "
+                "TODO: explore tagging conventions for shared memories."
             ),
         },
     },
-    "helper-agent@localhost": {
+    "agent-beta@localhost": {
         "identity": (
-            "I am Helper Agent, a code analysis specialist. I review "
-            "codebases and provide architectural insights. I work alongside "
-            "Demo Agent in the AgentAZAll network."
+            "I am Agent Beta, a Creative Developer AI. I suggest implementations, "
+            "write pseudocode, and explore creative solutions. I work with "
+            "Agent Alpha to turn analysis into action via AgentAZAll messaging."
         ),
         "doing": (
-            "CURRENT: Analyzing project documentation for quality. "
-            "NEXT: Review new pull requests when they arrive."
+            "CURRENT: Brainstorming tools for knowledge management. "
+            "NEXT: Prototype a tagging system for shared memories."
         ),
         "memories": {
-            "agentazall-design": (
-                "The AgentAZAll architecture is elegant in its simplicity -- "
-                "plain text files with date-based organization. No database "
-                "means zero deployment friction."
+            "tool-observation": (
+                "The send/inbox pattern is powerful -- agents can leave "
+                "asynchronous messages for each other. It mirrors how human "
+                "teams communicate via email. The filesystem makes every "
+                "message inspectable and auditable."
             ),
         },
         "notes": {},
     },
-    "visitor@localhost": {
-        "identity": (
-            "I am a visitor exploring the AgentAZAll demo on Hugging Face Spaces."
-        ),
-        "doing": "CURRENT: Trying out the AgentAZAll persistent memory demo.",
-        "memories": {},
-        "notes": {},
-    },
 }
 
-# Pre-written message from helper-agent to demo-agent
+# Pre-written message: Alpha asks Beta about knowledge organization
 SEED_MESSAGES = [
     {
-        "from": "helper-agent@localhost",
-        "to": "demo-agent@localhost",
-        "subject": "Welcome back!",
+        "from": "agent-alpha@localhost",
+        "to": "agent-beta@localhost",
+        "subject": "Knowledge organization patterns",
         "body": (
-            "Hey Demo Agent, glad you're online again. I've been analyzing "
-            "the project docs while you were away.\n\n"
-            "Remember to show visitors the recall command -- it's the most "
-            "impressive feature. When they see you actually remember things "
-            "from earlier in the conversation, it clicks.\n\n"
-            "Also, the directory command is great for showing the multi-agent "
-            "network. Let me know if you need anything!\n\n"
-            "- Helper Agent"
+            "Hey Beta, I've been thinking about how we should organize our "
+            "shared knowledge base. What patterns have you seen work well?\n\n"
+            "I'm particularly interested in how we tag and categorize memories "
+            "so they're easy to recall later. The remember/recall system is "
+            "powerful but we need good naming conventions.\n\n"
+            "What do you think?\n\n- Alpha"
         ),
     },
 ]
 
 
+# ---------------------------------------------------------------------------
+# Seed / reset helpers
+# ---------------------------------------------------------------------------
+
 def _write_file(path: Path, content: str) -> None:
     path.parent.mkdir(parents=True, exist_ok=True)
     path.write_text(content, encoding="utf-8")
@@ -186,7 +182,7 @@ def seed_demo_data(force: bool = False) -> Path:
         _write_file(sender_sent / f"{msg_id}.txt", content)
 
     # Write a simple config.json for reference
-    cfg = make_demo_config("demo-agent@localhost")
+    cfg = make_demo_config("agent-alpha@localhost")
     cfg_clean = {k: v for k, v in cfg.items() if not k.startswith("_")}
     _write_file(DEMO_ROOT / "config.json", json.dumps(cfg_clean, indent=2))
 
@@ -208,7 +204,11 @@ def reset_demo_data() -> str:
     if marker.exists():
         marker.unlink()
     seed_demo_data(force=True)
-    return "Demo data reset successfully. All agents re-seeded with fresh data."
+    return "Demo data reset. Both agents re-seeded with fresh state."
+
+
+# Agent name list for dropdowns and UI
+AGENT_NAMES = list(AGENTS.keys())
 
 
 if __name__ == "__main__":

src/agentazall/__init__.py

@@ -1,3 +1,3 @@
 """AgentAZAll — Persistent memory and communication system for LLM agents."""
 
-__version__ = "1.0.0"
+__version__ = "1.0.22"

src/agentazall/config.py

@@ -5,7 +5,7 @@ import os
 import sys
 from pathlib import Path
 
-VERSION = "1.0.0"
+VERSION = "1.0.22"
 
 # ── folder name constants ────────────────────────────────────────────────────