mcpField
M
memory-management

memory-management

github.com/jzocbAgent workflows39228 installsUpdated 38w ago
Repository

About memory-management

A skill for structuring agent memory in three layers, from always-loaded core notes to searchable lessons and raw daily logs. It defines priority-based retention periods and an archival routine for keeping active memory concise.

Key features

  1. 01Three-layer memory organization
  2. 02Priority tags and retention periods
  3. 03Searchable structured lesson files
  4. 04Automatic archival with a janitor script
  5. 05Memory size checks and dry runs

Use cases

  1. 01Keeping core agent notes concise
  2. 02Archiving expired project details
  3. 03Searching stored troubleshooting lessons
  4. 04Reviewing memory before cleanup

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/jzocb/openclaw-memory-management --skill memory-management

After installing, type/memory-managementin Claude Code to confirm it loaded.

Details

From the project README

Memory Management Skill

Manage agent memory efficiently using a three-layer architecture with automatic archival.

Quick Setup

  1. Format MEMORY.md with priority tags:
- [P0] Core identity item (never expires)
- [P1][2026-02-10] Active project (90-day TTL)
- [P2][2026-02-10] Temporary item (30-day TTL)
  1. Set up auto-archive cron:
0 4 * * * python3 ~/.openclaw/workspace/scripts/memory-janitor.py
  1. Store lessons in memory/lessons/*.jsonl:
{"id": "lesson-001", "date": "2026-02-10", "category": "infra", "title": "Problem title", "problem": "What happened", "solution": "How to fix", "tags": ["tag1"]}

Three-Layer Architecture

Layer 1: Hot Memory (MEMORY.md)

  • Always loaded, ≤200 lines
  • P0: Core identity, never expires
  • P1: Active projects, 90-day TTL
  • P2: Temporary, 30-day TTL

Layer 2: Cold Memory (searchable)

  • memory/lessons/*.jsonl — structured lessons
  • memory/archive/ — expired content
  • Use memory_search to recall

Layer 3: Raw Logs

  • memory/YYYY-MM-DD.md — daily logs
  • Not loaded automatically

Priority Guidelines

PriorityUse ForTTL
P0User identity, preferences, safety rulesNever
P1Active projects, current strategies90 days
P2Debug notes, one-time events30 days

Core Principles (max 5 in AGENTS.md)

Keep only essential rules in AGENTS.md. Other lessons go to lessons/*.jsonl.

Example 5 rules:

  1. Real money = correctness > speed
  2. External actions require confirmation
  3. Check both cron systems (system + OpenClaw)
  4. Long-running processes need setsid isolation
  5. Read platform rules before trading

When Memory Gets Too Large

Run janitor with stats:

python3 scripts/memory-janitor.py --stats

If >200 lines:

  1. Review P0 entries — are they all truly permanent?
  2. Add dates to P1/P2 entries missing them
  3. Move detailed content to lessons/*.jsonl
  4. Run --dry-run then archive

Comments

Markdown supported