신경 메모리 연구 저장소 안내

연구 기록 사본

원 저장소 · 원본 README.md · 수집 2026-10-16 · 커밋 fd190b18f8e4113d7f451a7108911c1f89716b24 당시의 실험 조건·수정·철회를 보존한 문서입니다. 성능 수치를 서로 다른 실험 간에 합산하지 않습니다. 연구 안내와 해석 범위부터 읽으세요. 코드·데이터·체크포인트는 이 사이트에 포함하지 않았습니다.

neural-memory-research

Whether a memory of past work makes an LLM agent better. Split out of the keymem repository so that research history and product history stop sharing a log.

Start with FINDINGS.md — what this repository establishes, what it does not, and the limit on each. The results files below are chronological and run to seven thousand lines between them; the findings file is the hundred-line version for someone who needs the conclusions rather than the path that produced them.

The short version: the central question is not answered. What is established is a family of approaches that does not work, and why — a memory built on judging whether the present resembles something seen before loses to a signal the model already emits for free. Explicit procedural knowledge is a separate case and does help, slightly.

Layout

• neural-memory/ — the agent experiments. The endpoint runs a coding agent on mutation-injected bugs and lets pytest decide, because scoring the likelihood of a recorded action cannot separate a better agent from a better predictor of one. Results in RESULTS.md, numbered by phase.
• fly-connectome/ — what the fly’s mushroom body wiring is organised for, measured against a rewiring that holds every degree fixed. Results in README.md, numbered by phase.
• board/ — a shared notice board from a stretch where two agent sessions worked this repository at once. claims.md holds what is believed and what measured it; 004-harness-hazards.md is the list of measurement defects found the hard way, and is probably more transferable than any result here.

Regenerating the data

Every corpus and derived feature is gitignored and has been deleted. Nothing here depends on them being present — the scripts rebuild them:

cd neural-memory && ./recover_artifacts.sh      # corpora, then the encoded streams
cd fly-connectome && uv run python fetch_mushroom_body.py

The Qwen encode is hours. Everything else is a download or minutes.

One thing worth knowing before reusing the harness

Task difficulty in the repair harness is almost entirely a property of which module a task mutates, and no measured property of a module predicts it — not size, not structure, not how many tests break. A pool drawn without stratifying by module produces condition effects that are module effects wearing condition labels. Phases 87 and 88.