mini-engram-hanhai

mini-engram demo adapters

Each directory is a "memory cartridge": pure LoRA weights (r=16, 40,370,176 params, bf16, ~81 MB), no base weights included. Attach one to Qwen/Qwen2.5-7B-Instruct and all of that tenant's baked-in facts come back.

The corpora are programmatically generated by scripts/make_demo_corpus.py and contain no real data — the base model cannot know these facts (closed-book F1 ≈ 0.12 confirms zero leakage).

Usage (vLLM multi-tenant serving)

# one process serving all tenant adapters, OpenAI-compatible API
vllm serve Qwen/Qwen2.5-7B-Instruct \
  --enable-lora --max-lora-rank 64 \
  --lora-modules xinglan=$PWD/xinglan hanhai=$PWD/hanhai

curl http://127.0.0.1:8000/v1/chat/completions -d '{
  "model": "xinglan", "messages": [{"role":"user","content":"What is your reimbursement system called?"}]
}'

Or use scripts/serve.sh + scripts/chat.py from the repository.

License

• These adapters are derivatives of Qwen/Qwen2.5-7B-Instruct (Apache 2.0); redistribution must retain its LICENSE and notices (see the Apache 2.0 summary attached in each directory's README.md).
• Our contribution is released under the mini-engram repository's MIT License, Copyright (c) 2026 TerseMind.

Citation

See docs/paper/ in the repository (paper "Bake It or Retrieve It", under arXiv submission).