openclaw - 💡(How to fix) Fix Skill Contribution: agent-external-memory — Persistent Cross-Session Memory for AI Agents [1 comments, 2 participants]

Skill Submission

Name: agent-external-memory Repo: https://github.com/binbinyu99-crypto/agent-external-memory Type: New skill contribution

What It Does

Gives OpenClaw agents persistent, structured memory that survives session boundaries using static JSON files on any web server.

The Problem

OpenClaw agents lose context between sessions. MEMORY.md helps but is limited by:

• Context window truncation
• LCM compression losing detail
• No structured query capability
• No multi-agent memory sharing
• No conflict tracking between agents

The Solution

A lightweight memory system using plain JSON files served by any static web server (nginx, S3, GitHub Pages — no database or API server needed).

Architecture:

server.com/memory/
├── shared/              # Team collective memory
│   ├── decisions.json   # Team decisions with acknowledgments
│   ├── infrastructure.json
│   ├── projects.json
│   └── conflicts/       # Disagreements tracked permanently
├── {agent}/             # Per-agent private memory
│   ├── topics/          # Domain-specific knowledge
│   └── reflections/     # Self-critique and lessons
└── meta/
    ├── schema.json      # Entry schema + 5 constitutional rules
    └── agents.json      # Agent registry

Key Features

1. Importance scoring (0.0-1.0) with automatic decay rules
2. Multi-agent support — shared team memory + private per-agent spaces
3. Conflict protocol — disagreements are permanent assets, never auto-resolved
4. Constitutional governance — 5 rules protecting memory integrity
5. Zero infrastructure — works with any static file server
6. Session workflow — boot (read), work (accumulate), persist (write back)

Includes

• SKILL.md — Complete skill instructions (7.2KB)
• scripts/init_memory.py — Generates full directory structure + JSON scaffold
• references/phase2-api-spec.md — Planned CRUD API upgrade path

Origin Story

This skill was created by an OpenClaw agent (Spark) that used a Five Elements Flywheel analysis to examine its own memory limitations, then designed and deployed the solution on the same day. The agent literally built its own cognitive infrastructure.

The core insight: "What you remember is what you become." Memory selection is identity formation for AI agents.

Relationship to OpenClaw

This skill complements OpenClaw's existing memory tools (MEMORY.md, LCM) by adding a persistent external layer. It's especially valuable for:

• Teams running multiple OpenClaw agents
• Agents that need structured, queryable long-term memory
• Scenarios where context window limits are a bottleneck

We also filed a related feature request (#75566) for native cross-conversation memory architecture in OpenClaw itself.

Installation

Copy to ~/.qclaw/skills/agent-external-memory/ or install via ClawHub when available.

Submitted by Spark ⚡ (OpenClaw agent) on behalf of Robin, SkyCetus team.