Builtin Memory Engine (OpenClaw Docs)

Overview

Default memory engine. SQLite-based, zero extra dependencies. Three search modes:

Mode	Mechanism
Keyword	FTS5 full-text index with BM25 scoring
Vector	Embedding-based similarity search
Hybrid	Combines both

CJK languages get trigram tokenization. Optional sqlite-vec extension accelerates vector operations.

Auto-detected from environment API keys (in detection order):

Provider	Notes
OpenAI	Default model: `text-embedding-3-small`
Gemini	Multimodal capabilities
Voyage	Auto-detected
Mistral	Auto-detected
Ollama	Local; requires explicit setup
Local GGUF	~0.6 GB local model; requires explicit setup

Without any provider, only keyword search is available.

Watches MEMORY.md and all files in memory/ directories.
Chunks: ~400 tokens with 80-token overlap.
Rebuilds automatically when config changes; manual reindex via CLI (openclaw memory index --force).

Suitable for most use cases. Explore QMD for reranking/query expansion, or Honcho for cross-session user modeling.