NativeMemoryProvider

Core, Recall and Wiki

The native provider uses PostgreSQL and pgvector directly. Core maintains pinned blocks; Recall projects recent turns with pagination for large messages; Wiki retrieves document chunks. Turn context is a bounded projection while the database preserves content and revisions.

pip install "orkmind[memory-provider]"
python -m orkmind.memory_provider.mcp

APIs and authorization

assemble_turn_context combines the layers. search_history requires a session or user; search_wiki and get_document share scope filtering. set_core_block is an operator API and must not be exposed as an agent tool. Missing required personas and invariants produce an error.

History and ingestion

The database protects history and revisions with append-only triggers. Ingestion normalizes content for deduplication, prepares chunks and embeddings, and writes the document transactionally. Incompatible vector dimension or search language changes are rejected. Chunks are derived and can be rebuilt; the full document remains the source.

Writing on behalf of a person

System ingestion has no owner. When a write comes from a person, pass their scope as writer: only what they could read is written. A classification out of reach raises ClassificationOutOfReachError before chunking or embedding; a document slug they cannot read raises SlugUnavailableError without revealing anything about the document that holds it. Both refusals inherit from WriteOutOfScopeError.

Content deduplication only considers what the scope reaches, and every update creates a new revision recorded in ingested_by. suggest_documents completes [[link]] only with what the scope sees.

scope = MemoryScopeFilter.for_user(user_context)
await memory.ingest_document(request, writer=scope)