agent-memory – skill Agent & Tự động hóa cho agent AI

Agent Memory — promotion is earned, not asserted

> Portability: stdlib only. No database, no embeddings, no network, no LLM calls.

The problem

A project's CLAUDE.md is a memory system with one tier and no eviction: every

durable fact and every passing preference land in the same always-loaded file,

until the important lines are diluted by the incidental ones. Facts learned

mid-session vanish at teardown unless someone writes them down.

The fix is not more storage — it is a promotion ladder. A claim earns its

way toward always-loaded context by recurring; a human confirms the last step.

The four tiers

Tiers are distinguished by injection policy, not storage format.

| Tier | Holds | Injected | Committed |

|---|---|---|---|

| L0 | raw session transcripts | never | no (already on disk) |

| L1 | candidate atoms | on relevance, at prompt time | no (gitignored) |

| L2 | this project's context | every session start | yes, after adopt |

| L3 | stable cross-project persona | always | yes, after adopt |

The gates

Nothing moves up because it sounded important. It moves up because it recurred.

• L0 → L1 — an explicit marker fires (a directive, a correction, a stated

preference, a named lesson, a reproducible failure). Rule-based, high

precision, deliberately low recall.

• L1 → L2 — ≥ 3 distinct sessions spanning ≥ 2 distinct calendar days. A

claim stated outright needs 2 sessions; the distinct-day rule still applies. A

verified claim promotes on one observation and is the only day-exempt path.

• L2 → L3 — held in ≥ 2 distinct projects, aged ≥ 30 days, uncontested.

Two gates refuse rather than guess. A claim whose text was altered by

redaction never promotes on evidence alone — the flag firing is evidence the

source was sensitive, and a lexical filter finding one secret is not proof it

found all of them. A claim with an open contradiction is frozen at L1 until a

human resolves it; the incumbent is never silently overwritten.

Use it

# what is remembered, and what is blocking the next promotion
python3 scripts/memory_inspect.py --tier L1

# where did this line come from — sessions, days, transcript, quoted source
python3 scripts/memory_inspect.py --why "PR base branch is dev"

# every claim with an open contradiction, both directions of the join
python3 scripts/memory_inspect.py --contested

# dry-run the promotion pass; writes nothing
python3 scripts/memory_promote.py

Three hooks run the loop unattended: SessionStart injects L2 + L3,

UserPromptSubmit recalls relevant L1 atoms, SessionEnd captures and stages.

Each is disabled independently with AGENT_MEMORY_SESSIONSTART=0,

AGENT_MEMORY_USERPROMPTSUBMIT=0, AGENT_MEMORY_SESSIONEND=0. Every hook fails

open: a broken memory system costs you memory, never a session.

Hard rules

1. Redact before writing. Every atom passes the filter before it reaches

disk. Anything altered is quarantined from promotion.

2. Propose, never apply. Promotions land in .memory/staged/. Only an

explicit /cs-memory adopt touches a CLAUDE.md, and it backs both up first.

3. Cite, don't invent. Every atom carries a back-pointer to the transcript

line that produced it. --why resolving to *ambiguous* prints nothing rather

than guess: a wrong citation is worse than a missing one.

4. Never surface a contested claim as fact. It is still injected — hiding

the conflict is worse — but always tagged.

5. The committed tiers carry no paths. Promotion strips the back-pointer

prefix, which embeds an OS username.

Forcing questions

Walk these one at a time before trusting the store.

1. Which line in your CLAUDE.md did you last actually read before acting?

2. Would you rather the agent forget a true thing, or remember a false one?

3. When two remembered rules disagree, who decides — and when?

4. What would make you delete .memory/ entirely?

Rationale, open decisions, field schema: [../../DESIGN.md](../../DESIGN.md).