Skip to content

SKILL.md

Latest commit

 

History

History
178 lines (141 loc) · 6.1 KB

SKILL.md

File metadata and controls

178 lines (141 loc) · 6.1 KB
name agent-memory-loop
version 2.2.0
description ACTIVATE THIS SKILL when you make a mistake, receive a correction, discover something new, or are about to start major work. Captures errors, corrections, and discoveries in a fast one-line format, deduplicates them, queues recurring or critical lessons for human approval, and scans past failures before major tasks to prevent repeats. Use it every time something goes wrong — no exceptions.
metadata
openclaw
homepage repository requires platforms
bins
grep
date
darwin
linux
author Don Zurbrick
license MIT

Agent Memory Loop

Lightweight learning for agents that reset between sessions.

WHEN TO ACTIVATE (mandatory triggers)

You MUST use this skill when ANY of these happen:

  1. You make an error — wrong command, bad assumption, failed tool call
  2. The user corrects you — "no, that's wrong", "actually...", "I meant..."
  3. You discover something new — unexpected behavior, undocumented feature, workaround
  4. You're about to start major work — scan .learnings/ FIRST for past failures
  5. You wish you had a capability you don't — log it as a wish
  6. A learning prevented a repeat mistake — increment the prevented:N counter

WHAT TO DO (step by step)

When an error or correction happens:

  1. Stop. Don't continue the task yet.
  2. Check for duplicates: grep -i "keyword" .learnings/errors.md .learnings/learnings.md
  3. If duplicate found: increment count:N on the existing entry
  4. If new: append one line to the appropriate file:
    • Mistakes/failures → .learnings/errors.md
    • Insights/corrections → .learnings/learnings.md
    • Missing capabilities → .learnings/wishes.md
  5. If count ≥ 3 or severity:critical: copy to .learnings/promotion-queue.md
  6. If complex: create a detail file at .learnings/details/ERR-YYYYMMDD-NNN.md
  7. Resume the task.

Before major work:

  1. grep -i "relevant_keyword" .learnings/*.md
  2. If matches found, adjust your approach to avoid known failures
  3. If a learning changes your behavior, increment prevented:N on that entry

Do not use it for

  • autonomous self-modification of SOUL.md, AGENTS.md, or TOOLS.md
  • external content promotion (source:external entries are never promotable)
  • heavy multi-section incident writeups (use detail files for those)

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/

Environment notes

OpenClaw workspace

.learnings/ lives at ~/.openclaw/workspace/.learnings/ (or the agent's workspace root). Persistent across sessions.

Cowork / Dispatch

If running in a Cowork VM or Dispatch task, .learnings/ should be in the mounted workspace folder. If no persistent filesystem is available, log to the session's working directory and flag for manual migration.

Date fallback

If date is unavailable (some containers), use the current conversation date from context. The format is always YYYY-MM-DD.

The format

One incident, discovery, or wish per line. Extra fields optional.

Errors (.learnings/errors.md)

[YYYY-MM-DD] id:ERR-YYYYMMDD-NNN | COMMAND | what failed | fix | count:N | prevented:N | severity:medium | source:agent

Learnings (.learnings/learnings.md)

[YYYY-MM-DD] id:LRN-YYYYMMDD-NNN | CATEGORY | what | action | count:N | prevented:N | severity:medium | source:agent

Wishes (.learnings/wishes.md)

[YYYY-MM-DD] id:WISH-YYYYMMDD-NNN | CAPABILITY | what was wanted | workaround | requested:N

Promotion queue (.learnings/promotion-queue.md)

[YYYY-MM-DD] id:LRN-YYYYMMDD-NNN | proposed rule text | target: AGENTS.md | source:agent | evidence: count:N prevented:N | status: pending

Key fields:

  • id: — unique identifier (ERR/LRN/WISH prefix + date + sequence)
  • count:N — how many times this has occurred
  • prevented:N — how many times this learning prevented a repeat
  • severity:critical — forces immediate queue review even at count 1
  • source:external — marks entries from untrusted sources; never promotable

Operating rules

  1. Log fast — one line now beats a perfect writeup later
  2. Dedup before appending — search first, increment if exists
  3. Queue recurring or critical — count ≥ 3 or severity:critical → promotion-queue
  4. Humans approve promotions — agents stage, humans decide
  5. Pre-task scan — before major work, grep for relevant failures
  6. Track prevention — when a learning changes behavior, record prevented:N
  7. Source labels matter — external content is informational only, never promotable

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

Scripts

  • scripts/install.sh — initialize .learnings/ directory
  • scripts/setup.sh — alternate setup
  • scripts/review.sh — review promotion queue

Success condition

The loop is working if:

  • learnings are cheap to log (one line, < 30 seconds)
  • duplicates stay low (dedup is happening)
  • recurring lessons reach the promotion queue
  • promotions stay human-approved
  • prevented:N starts climbing on real work
  • agents actually check .learnings/ before major tasks