Memory Pipeline

Every meaningful exchange gets summarized, embedded, indexed, and made searchable across 44 per-agent SQLite databases. Search runs at 70% vector + 30% BM25 keyword hybrid.

The components

• Inputs: anything that flows through Claude — conversation text, file contents, tool results — gets queued for embedding.
• Embedding: Voyage-4 via Portkey (config slug pc-opencl-22d508). Embeddings are cached by content hash so duplicate text doesn’t re-embed.
• Storage: 44 SQLite databases at ~/.openclaw/memory/<agent>.sqlite. Each DB has 5 tables:
  • chunks — raw text + metadata
  • chunks_fts — SQLite FTS5 keyword index
  • chunks_vec — vec0 vector index
  • pending_saves — queue for async writes
  • dead_letters — failed embeddings for retry
• Hybrid retrieval: when you call memory_search, both indexes run in parallel. Vector search weights 0.7 (semantic similarity); BM25 keyword weights 0.3 (exact match). Scores fuse and rerank.

Per-agent isolation

Each agent has its own DB. The aurora agent’s memory is separate from acq’s. To search across all 44, the memory CLI accepts --agent <name> to scope or scan multiple.

How to query it

python3 ~/.openclaw/tools/claude-memory.py search "your query"
python3 ~/.openclaw/tools/claude-memory.py search "your query" --agent acq
python3 ~/.openclaw/tools/claude-memory.py stats

Or via MCP from inside Claude: memory_search (openclaw-tools server) returns top-K hits.

Total scale

~5,568 chunks fleet-wide as of last fleet check. Largest DBs are agent-specific working sets (aurora, acq, dispo); smallest are fresh agents (single-digit chunks).