agent-memory-loop（Agent 记忆循环）

Agent Memory Loop

Lightweight learning for agents that reset between sessions.

Use this when

• you want a low-friction way to log mistakes, corrections, and discoveries
• you need recurring lessons without bloating core instructions
• you want human-reviewed promotion instead of auto-writing to instruction files
• you want a quick pre-task scan for known failure patterns

Do not use it for

• autonomous self-modification
• external content promotion
• heavy multi-section incident writeups by default
• dashboards, registries, or process ceremony

Core workflow

error / correction / discovery
        ↓
log one line in .learnings/
        ↓
dedup by id, then keyword
        ↓
count:3+ or severity:critical → promotion-queue
        ↓
human reviews promotion
        ↓
check relevant learnings before major work
        ↓
increment prevented:N when a learning actually changed behavior

Install

bash scripts/install.sh

Creates:

.learnings/
  errors.md
  learnings.md
  wishes.md
  promotion-queue.md
  details/
  archive/

Minimal instruction snippet

Add this to your agent instructions:

## Self-Improvement
Before major tasks: grep .learnings/*.md for relevant past issues.
After errors or corrections: log a one-line entry using agent-memory-loop.
Never auto-write to SOUL.md, AGENTS.md, TOOLS.md, or similar instruction files.
Stage candidate rule changes in .learnings/promotion-queue.md for human review.

The format, in short

One incident or discovery per line. Extra fields are optional.

[YYYY-MM-DD] id:ERR-YYYYMMDD-NNN | COMMAND | what failed | fix | count:N | prevented:N | severity:medium | source:agent
[YYYY-MM-DD] id:LRN-YYYYMMDD-NNN | CATEGORY | what | action | count:N | prevented:N | severity:medium | source:agent
[YYYY-MM-DD] CAPABILITY | what was wanted | workaround | requested:N
[YYYY-MM-DD] id:LRN-YYYYMMDD-NNN | proposed rule text | target: AGENTS.md | source:agent | evidence: count:N prevented:N | status: pending

Key fields:

• count:N tracks recurrence
• prevented:N tracks loop closure
• severity:critical forces review even at count 1
• source:external is never promotable

Operating rules

1. Log fast; prefer a one-line entry over a perfect writeup
2. Dedup before appending
3. Queue recurring or critical lessons for review
4. Humans approve promotions; agents do not
5. Before major work, scan for relevant prior failures
6. If a learning prevented a repeat mistake, record that with prevented:N

References

• references/logging-format.md — canonical line formats, fields, examples, source labels
• references/operating-rules.md — dedup, review queue, pre-task review, trimming rules
• references/promotion-queue-format.md — queue entry structure and status lifecycle
• references/detail-template.md — optional detail-file template for complex failures
• references/design-tradeoffs.md — why this stays lean instead of turning into a system

Assets and scripts

• assets/errors.md
• assets/learnings.md
• assets/wishes.md
• assets/promotion-queue.md
• scripts/install.sh
• scripts/setup.sh
• scripts/review.sh

Success condition

The loop is working if agents actually use it:

• learnings are cheap to log
• duplicates stay low
• recurring lessons reach the queue
• promotions stay human-approved
• prevented:N starts climbing on real work