/agent-memory

name: agent-memory
description: "Use when an agent keeps relearning the same correction, CLAUDE.md has grown unreadable, or you need to trace where a remembered rule came from. Promotes facts through four tiers only on recurrence across sessions and days, with memory_promote.py and a human adopt step."
argument-hint: "[optional: status | why \"<claim>\" | a tier name]"
license: MIT
metadata:
  version: 1.0.0
  build_pattern: "Tencent TencentDB-Agent-Memory's tiering concept rebuilt natively on Claude Code hooks; deterministic recurrence gates, no LLM, no database"
  distinct_from: "llm-wiki (a vault you write on purpose; this writes itself from sessions); skillopt-sleep (replays tasks to improve a skill; this extracts facts to remember); memory-engineering (audits and prices any memory system; this IS one, and is a legitimate subject of that audit)"

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).

More in this skill

Load these only when the task needs them.

• [references/hook_capture_discipline.md](references/hook_capture_discipline.md) — Hook discipline — running on someone else's critical path
• [references/promotion_gate_design.md](references/promotion_gate_design.md) — Designing the gates — why these thresholds, and what they cost
• [references/tiered_memory_canon.md](references/tiered_memory_canon.md) — Tiered agent memory — the canon, and what this skill takes from it

• `python3 scripts/memory_extract.py --help` — L0 -> L1. Rule-based, no LLM
• `python3 scripts/validate_examples.py` — drift gate for DESIGN.md, memory_schema.json and the worked examples