🧠 Agent Memory Loop

Lightweight self-improvement loop for AI agents.

Your agent forgets everything between sessions. This skill gives it a learning system that actually works — one-line entries, structured dedup, severity-aware review queues, injection-safe source labels, and pre-task review. Minimal context burn, maximum learning.

Quick Start

bash scripts/install.sh /path/to/workspace

Then add to your agent's instructions:

## Self-Improvement
Before major tasks: `grep -i "keyword" .learnings/*.md` for relevant past issues.
After errors or corrections: log to `.learnings/` using the agent-memory-loop format.
Never auto-write to SOUL.md/AGENTS.md/TOOLS.md. Stage to .learnings/promotion-queue.md.

How It Works

1. Log errors, corrections, and discoveries as one-line entries
2. Dedup by stable ID (fallback: keyword grep)
3. Review queue when recurring (count:3+) or critical (severity:critical)
4. Human approves promotion to instruction files
5. Pre-task review before major work — grep, name the learning, state the adjustment
6. Track prevention — increment prevented:N when a learning actually changed behavior

Runtime structure

File	Purpose
SKILL.md	Lean runtime entrypoint
references/logging-format.md	Canonical line formats, optional fields, examples
references/operating-rules.md	Dedup, review queue, promotion model, trimming
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/install.sh	Set up .learnings/ in a workspace
scripts/review.sh	Health check — pending promotions, stale entries, stats
assets/*.md	Template files copied by install script

Key features

• Review queue — no auto-promotion to instruction files; human approval required
• Source labels — agent / user / external; external content blocked from promotion
• Severity awareness — severity:critical triggers review even at count:1
• Loop closure — prevented:N tracks whether learnings actually changed behavior
• Structured dedup — stable IDs (ERR-YYYYMMDD-NNN) instead of raw grep
• Optional detail files — link to .learnings/details/ for complex failures
• Staleness / expiry — optional expires: field + periodic trimming

Requirements

• grep, date (any POSIX system)
• No frameworks, no dependencies, no configuration

License

MIT — Don Zurbrick

Links

• ClawHub
• GitHub