Openclaw Memories

Agent memory with ALMA meta-learning, LLM fact extraction, and full-text search. Observer calls remote LLM APIs (OpenAI/Anthropic/Gemini). ALMA and Indexer w...

OpenClaw Memory System

Three components for agent memory:

1. ALMA — Evolves memory designs through mutation + evaluation (offline)
2. Observer — Extracts structured facts from conversations via LLM API (requires API key)
3. Indexer — Full-text search over workspace Markdown files (offline)

Environment Variables

Observer requires one of:

• OPENAI_API_KEY
• ANTHROPIC_API_KEY
• Or pass apiKey in config

ALMA and Indexer require no keys or network access.

How It Works

ALMA (Algorithm Learning via Meta-learning Agents)

Proposes memory system designs, evaluates them, keeps the best. Uses gaussian mutation and simulated annealing to explore the design space.

alma.propose() → design
alma.evaluate(design.id, metrics) → score  
alma.best() → top design
alma.top(5) → leaderboard

Observer

Sends conversation history to an LLM, gets back structured facts:

• Kind: world fact / biographical / opinion / observation
• Priority: high / medium / low
• Entities: mentioned people/places
• Confidence: 0.0–1.0 for opinions

Fails gracefully — returns empty array if LLM is unavailable.

Indexer

Chunks workspace Markdown files and indexes them for search:

• MEMORY.md — core facts
• memory/YYYY-MM-DD.md — daily logs
• bank/entities/*.md — entity summaries
• bank/opinions.md — beliefs with confidence

indexer.index() → count of chunks indexed
indexer.search('query') → ranked results
indexer.rebuild() → re-index from scratch

Install

npm install @artale/openclaw-memory

Limitations

• Indexer uses an in-memory mock database, not real SQLite FTS5. Search works but ranking is simplified.
• Observer calls remote APIs — not offline. Only ALMA and Indexer work without network.
• No dashboard — removed in v2 for simplicity.

Source

5 files, 578 lines, 0 runtime dependencies.

https://github.com/arosstale/openclaw-memory