openclaw - 💡(How to fix) Fix Support PostgreSQL/pgvector as a memory search store driver

Feature Request

Problem

The memory search system currently only supports SQLite as its store driver (agents.defaults.memorySearch.store.driver is const: "sqlite"). For deployments that already run PostgreSQL with pgvector for other embedding/vector search needs, this creates two parallel vector systems that don't share data:

1. OpenClaw memory search (SQLite) — indexes memory files and session transcripts
2. Application-level embeddings (PostgreSQL/pgvector) — entity facts, lessons, events, research, etc.

The memory_search tool only queries the SQLite index, missing all the PostgreSQL-stored knowledge. Meanwhile, custom hooks/plugins that query PostgreSQL can't benefit from OpenClaw's hybrid BM25+vector search pipeline.

Proposed Solution

Add PostgreSQL/pgvector as a supported store.driver option alongside SQLite. This would allow deployments to:

• Unify all vector embeddings in a single PostgreSQL database
• Query both memory files and application data through the memory_search tool
• Leverage PostgreSQL's mature backup, replication, and monitoring infrastructure
• Use pgvector's HNSW/IVFFlat indexes for larger-scale deployments

Alternatives Considered

• Markdown export bridge: Periodically dump PostgreSQL data into markdown files for OpenClaw to index. Works but creates stale snapshots and doubles storage.
• extraPaths with generated files: Similar, uses memorySearch.extraPaths to point at exported markdown. Same staleness issue.
• Custom turn-context plugins: Inject PostgreSQL data per-turn. Works for proactive recall but doesn't surface through the memory_search tool when the agent explicitly searches.

Context

Running OpenClaw with a PostgreSQL database containing ~15,000 semantic embeddings (entity facts, lessons, events, research) via pgvector. The memory search SQLite index has ~1,100 chunks from memory files and session transcripts. Having both systems queryable through a single memory_search call would significantly improve recall quality.

Environment

• OpenClaw version: 2026.5.7
• PostgreSQL 16 with pgvector
• Current memory search provider: Ollama (mxbai-embed-large)