mind-mem-4b (v3.0.0)

A governance-aware memory-assistant model for mind-mem — an auditable, contradiction-safe memory layer for coding agents (MCP-compatible).

This checkpoint is a full fine-tune of Qwen/Qwen3.5-4B, fine-tuned on the mind-mem 3.0.0 source tree: all 57 MCP tool signatures, 14 block-type schemas, full CHANGELOG history through v3.0.0, the docs/ tree, and a curated set of end-to-end governance workflow transcripts.

What it knows about

• 57 MCP tools — exact signatures, arg types, return envelopes, scope requirements.
• 14 block schemas — ADR, CODE, PERF, ALGO, BUG, DEC, CONV, DREF, CHECK, EV, FIELD, TIER, IMAGE, AUDIO.
• Governance workflows — propose → list_contradictions → approve_apply → verify_chain → rollback with BeliefStore + FieldAuditor + AuditChain wiring.
• Drift detection — live DriftDetector semantic pass alongside the lexical DRIFT.md pass (v3.0.0+).
• Memory tiers — 4-tier promotion cycle (WORKING → SHARED → LONG_TERM → VERIFIED), tier-boost retrieval ranking.
• Encryption — admin-scope encrypt_file / decrypt_file MCP tools gated on MIND_MEM_ENCRYPTION_PASSPHRASE.

Usage

Load the model (bf16)

from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL = "star-ga/mind-mem-4b"

tokenizer = AutoTokenizer.from_pretrained(MODEL)
model = AutoModelForCausalLM.from_pretrained(MODEL, torch_dtype="bfloat16", device_map="auto")

messages = [
    {"role": "system", "content": "You are mind-mem-4b, a memory-governance assistant."},
    {"role": "user",   "content": "Which MCP tool should I call to verify my audit chain?"},
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

Quantized (GGUF) inference with llama.cpp

# Grab the Q4_K_M build
huggingface-cli download star-ga/mind-mem-4b mind-mem-4b-Q4_K_M.gguf --local-dir ./gguf

# Run via llama-server, llama-cli, Ollama, LM Studio …
llama-cli -m ./gguf/mind-mem-4b-Q4_K_M.gguf -p "Show me a DREF block template."

Training recipe

Knob	Value
Base	Qwen/Qwen3.5-4B
Method	Full fine-tune (bf16, all 4.2B params trained, AdamW 8-bit)
LoRA rank	16
LoRA alpha	32
LoRA dropout	0.05
Target modules	q,k,v,o,gate,up,down-proj (all linear)
Epochs	3
Per-device batch size	1
Gradient accumulation	16
Learning rate	2e-4 (cosine, 3% warmup)
Precision	bf16
Optimizer	paged AdamW 8-bit
Hardware	RTX 3080 10GB

Corpus

Built deterministically from the mind-mem 3.0.0 source tree. Running python3 train/build_corpus.py in the repo reproduces the exact training JSONL byte-for-byte. Five disjoint sources:

1. MCP tool docstrings (src/mind_mem/mcp_server.py — 57 tools)
2. Block-type schemas (14 templates + field lists)
3. CHANGELOG entries (v1.0.0 → v3.0.0)
4. docs/ prose (setup, usage, api-reference, architecture, roadmap)
5. Curated governance workflow transcripts (6 scenarios)

All five sources are local to the repo — no external LLM calls, no web scraping, no synthetic data from a teacher model. The training data is auditable.

Eval

Three held-out benchmarks, scored zero-shot on the adapter-loaded base. See train/eval_harness.py for the exact harness — it runs on every commit to catch regressions.

Benchmark	Score	Items
Tool-call name recall	65%	13/20 prompts cite the correct MCP tool name
Block-schema conformance	70%	7/10 templates include canonical field names + ID prefix
Governance workflow	60%	3/5 workflows respond with the correct tool chain

Honest read

These numbers are an improvement over the prior adapter (v2.8.x on Qwen3.5-4B), which was trained pre-v2 API and does not know about any of the 35 new MCP tools shipped between v1.9 → v2.9 (it scores 0% on every v2.x-specific prompt). But they're below the aspirational 95 / 98 / 90% gates; the gap is concentrated in three failure modes:

1. Imperative phrasing ("Roll back an apply.") still occasionally triggers role-play responses instead of tool recall — Full-FT at 1,450 examples × 5 epochs still leaves residual base-model priors on imperative phrasings. v3.1 will expand the corpus to 10k+ examples.
2. Block-template hallucination — the model sometimes invents plausible-sounding fields instead of the canonical ones (e.g. inventing EvidenceType: where the schema requires Signal:).
3. Workflow-as-prose — "Walk me through" prompts sometimes produce explanatory prose instead of a tool chain.

Future iterations will address these with (a) a 3-5k-example corpus including more diverse imperative phrasings, (b) schema-conformance reinforcement with negative examples, and (c) rank-64 LoRA.

Use the base Qwen/Qwen3.5-4B plus this adapter when you want mind-mem-aware answers; use the base model alone for open-domain chat.

Intended use / scope

This is a specialised assistant, not a general-purpose LLM. It's tuned to answer questions about mind-mem internals, help agents compose correct MCP calls, and narrate governance workflows. Use the base Qwen3.5-4B for open-domain chat.

License

Apache-2.0 (same as the mind-mem Python package).

Changelog

• v3.0.0 (2026-04-14): Full fine-tune on Qwen/Qwen3.5-4B — all 4.2B parameters trained (not LoRA). Final loss 0.65, token accuracy 0.86. Covers mind-mem v1.9 → v3.0 surface: 57 MCP tools, 14 block schemas, governance workflows (evidence chain, field audit, drift, tier decay, alerting hooks, transparent encryption).
• v2.9.0: Legacy QLoRA adapter on Qwen/Qwen2.5-7B-Instruct base. Superseded by v3.0.0.
• v2.8.x: Initial release on Qwen3.5-4B base.

Citation

@software{mind_mem_7b_2026,
  author = {STARGA, Inc.},
  title = {mind-mem-4b: governance-aware memory-assistant for coding agents},
  year = 2026,
  version = {v3.0.0},
  url = {https://huggingface.co/star-ga/mind-mem-4b}
}