Skip to content
Canonical markdown: architecture/OPENCLAW_MEMORY_ARCHITECTURE.md. This rendered HTML page is supplemental reference only.

OpenClaw Memory Architecture

Last Updated: 2026-03-12

Current Memory Model

  • The bridge/session path determines the effective memory scope for a conversation.
  • OpenClaw memory-search is a separate configuration lane from OpenClaw chat generation.
  • The current memory-search provider is gemini with model gemini-embedding-001.
  • Memory-search fallback is treated separately from the chat fallback chain.

Document RAG Embedding Is A Separate Pipeline

  • The document ingestion pipeline (PDF, DOCX, PPTX, JSON, HTML, Google Docs/Drive) uses its own experimental embedding lane for semantic indexing into pgvector. The specific model and dimensionality are determined by a bounded evaluation spike and are not yet canonical default policy.
  • This is distinct from the OpenClaw memory-search lane above, which uses gemini-embedding-001 (768 dimensions) for conversational memory retrieval.
  • The two pipelines share the pgvector storage layer but use separate vector stores.

Why This Matters

  • It prevents confusion between chat generation models, memory-search models, and document RAG models.
  • It keeps scope and namespace behavior explicit.
  • It explains why memory configuration changes are not the same as chat-model or document-embedding changes.