File size: 10,350 Bytes
22026ba | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 | """
CoALA-inspired Memory Architecture for AETHER.
Four modules: Working, Episodic, Semantic, Procedural.
Plus Temporal Memory for long-horizon reasoning.
"""
import torch
import torch.nn as nn
from collections import deque
from typing import Dict, List, Any, Optional
import time
import json
class WorkingMemory:
"""
Active scratchpad for current reasoning cycle.
Limited capacity, fast access.
"""
def __init__(self, capacity: int = 16):
self.capacity = capacity
self.buffer: deque = deque(maxlen=capacity)
self.attention_weights = nn.Parameter(torch.ones(capacity))
def store(self, item: Dict[str, Any]):
item["_timestamp"] = time.time()
self.buffer.append(item)
def retrieve(self, query: str, top_k: int = 3) -> List[Dict]:
"""Simple attention-based retrieval from working buffer."""
if not self.buffer:
return []
# Compute relevance scores (simplified)
scores = []
for i, item in enumerate(self.buffer):
score = sum(1 for k in item if query.lower() in str(k).lower())
scores.append(score * torch.sigmoid(self.attention_weights[i]).item())
# Get top-k indices
indices = sorted(range(len(scores)), key=lambda i: scores[i], reverse=True)[:top_k]
return [list(self.buffer)[i] for i in indices]
def export(self) -> List[Dict]:
return list(self.buffer)
def __len__(self):
return len(self.buffer)
class EpisodicMemory:
"""
Experience buffer storing past interactions.
Temporal structure for long-horizon reasoning.
"""
def __init__(self, buffer_size: int = 1000):
self.buffer_size = buffer_size
self.buffer: deque = deque(maxlen=buffer_size)
def store(self, episode: Dict[str, Any]):
episode["_timestamp"] = time.time()
self.buffer.append(episode)
def retrieve_similar(self, query: str, top_k: int = 5) -> List[Dict]:
"""Retrieve episodes similar to query."""
if not self.buffer:
return []
# Simple keyword matching (replace with embedding-based in production)
scores = []
for item in self.buffer:
text = json.dumps(item)
score = sum(1 for word in query.lower().split() if word in text.lower())
scores.append(score)
indices = sorted(range(len(scores)), key=lambda i: scores[i], reverse=True)[:top_k]
return [list(self.buffer)[i] for i in indices]
def get_recent(self, n: int = 10) -> List[Dict]:
"""Get n most recent episodes."""
return list(self.buffer)[-n:]
def export(self) -> List[Dict]:
return list(self.buffer)
def __len__(self):
return len(self.buffer)
class SemanticMemory:
"""
World knowledge - external and learned facts.
Backed by knowledge graph (see knowledge.py).
"""
def __init__(self):
self.facts: Dict[str, Any] = {}
def store_fact(self, key: str, value: Any, confidence: float = 1.0):
self.facts[key] = {"value": value, "confidence": confidence, "timestamp": time.time()}
def retrieve(self, key: str) -> Optional[Dict]:
return self.facts.get(key)
def query(self, query: str) -> List[Dict]:
"""Simple prefix matching."""
return [v for k, v in self.facts.items() if query.lower() in k.lower()]
def export(self) -> Dict:
return self.facts
class ProceduralMemory:
"""
Learned skills, tool definitions, and code implementations.
Inspired by Yunjue Agent's tool accumulation.
"""
def __init__(self):
self.tools: Dict[str, Dict] = {}
self.tool_usage_stats: Dict[str, int] = {}
def register_tool(self, name: str, code: str, description: str,
tags: List[str] = None):
self.tools[name] = {
"code": code,
"description": description,
"tags": tags or [],
"registered_at": time.time(),
"version": 1,
}
self.tool_usage_stats[name] = 0
def get_tool(self, name: str) -> Optional[Dict]:
if name in self.tools:
self.tool_usage_stats[name] += 1
return self.tools[name]
return None
def search_tools(self, query: str) -> List[Dict]:
"""Search tools by description or tags."""
results = []
for name, tool in self.tools.items():
text = f"{name} {tool['description']} {' '.join(tool['tags'])}"
if query.lower() in text.lower():
results.append({"name": name, **tool})
return results
def merge_tools(self, tool_cluster: List[str]) -> Optional[str]:
"""
Merge functionally redundant tools (Yunjue-style tool absorption).
Returns name of merged tool or None.
"""
if len(tool_cluster) < 2:
return None
# Simple merge: keep highest usage tool as canonical
canonical = max(tool_cluster, key=lambda t: self.tool_usage_stats.get(t, 0))
# Merge descriptions
merged_desc = " | ".join(
self.tools[t]["description"] for t in tool_cluster if t in self.tools
)
self.tools[canonical]["description"] = merged_desc
self.tools[canonical]["version"] += 1
# Remove redundant tools
for t in tool_cluster:
if t != canonical and t in self.tools:
del self.tools[t]
return canonical
def export(self) -> Dict:
return {
"tools": self.tools,
"usage_stats": self.tool_usage_stats,
}
class CoALAMemory:
"""
Unified memory system following CoALA cognitive architecture.
Combines Working, Episodic, Semantic, and Procedural memory.
"""
def __init__(self, capacity: int = 16):
self.working = WorkingMemory(capacity=capacity)
self.episodic = EpisodicMemory(buffer_size=1000)
self.semantic = SemanticMemory()
self.procedural = ProceduralMemory()
def store(self, item: Dict[str, Any], memory_type: str = "working"):
if memory_type == "working":
self.working.store(item)
elif memory_type == "episodic":
self.episodic.store(item)
elif memory_type == "semantic":
for k, v in item.items():
self.semantic.store_fact(k, v)
elif memory_type == "procedural":
if "name" in item and "code" in item:
self.procedural.register_tool(
item["name"], item["code"],
item.get("description", ""),
item.get("tags", [])
)
def retrieve(self, query: str, memory_type: str = "all", top_k: int = 5) -> List[Dict]:
if memory_type == "all":
results = []
results.extend(self.working.retrieve(query, top_k=top_k//2))
results.extend(self.episodic.retrieve_similar(query, top_k=top_k))
results.extend(self.semantic.query(query)[:top_k])
return results[:top_k]
elif memory_type == "working":
return self.working.retrieve(query, top_k)
elif memory_type == "episodic":
return self.episodic.retrieve_similar(query, top_k)
elif memory_type == "semantic":
return self.semantic.query(query)[:top_k]
elif memory_type == "procedural":
return self.procedural.search_tools(query)
return []
@property
def buffer(self):
"""Alias for working memory buffer."""
return self.working.buffer
def export(self) -> Dict[str, Any]:
return {
"working": self.working.export(),
"episodic": self.episodic.export(),
"semantic": self.semantic.export(),
"procedural": self.procedural.export(),
}
class TemporalMemory(nn.Module):
"""
Time-sensitive memory with learned temporal attention.
Enables long-horizon reasoning and contextual adaptation.
Uses a simple LSTM-like gating mechanism.
"""
def __init__(self, buffer_size: int = 1000, hidden_dim: int = 64):
super().__init__()
self.buffer_size = buffer_size
self.hidden_dim = hidden_dim
self.buffer: deque = deque(maxlen=buffer_size)
# Temporal attention network
self.temporal_gate = nn.Sequential(
nn.Linear(2, hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim, 1),
nn.Sigmoid(),
)
def store(self, event: Dict[str, Any]):
event["_timestamp"] = time.time()
self.buffer.append(event)
def retrieve_context(self, current_time: Optional[float] = None,
lookback_window: float = 3600.0) -> List[Dict]:
"""
Retrieve events within lookback window, weighted by recency.
"""
current_time = current_time or time.time()
relevant = []
for event in self.buffer:
age = current_time - event.get("_timestamp", current_time)
if age <= lookback_window:
# Temporal relevance score: exponential decay
recency_score = torch.exp(torch.tensor(-age / lookback_window))
relevant.append({
**event,
"recency_score": recency_score.item(),
"age_seconds": age,
})
# Sort by recency score
relevant.sort(key=lambda x: x["recency_score"], reverse=True)
return relevant
def retrieve_with_attention(self, query_embedding: torch.Tensor,
top_k: int = 10) -> List[Dict]:
"""
Attention-based retrieval combining temporal and semantic relevance.
(Placeholder - would use actual embeddings in full implementation)
"""
return self.retrieve_context()[:top_k]
def export(self) -> List[Dict]:
return list(self.buffer)
@property
def buffer_contents(self):
return list(self.buffer)
def __len__(self):
return len(self.buffer)
|