strands-postgres-store

A PostgreSQL + pgvector MemoryStore for Strands Agents — semantic long-term agent memory backed by native pgvector similarity search (the <=> cosine-distance operator over an HNSW index). No external vector database.

pip install strands-postgres-store

from strands import Agent
from strands.memory import MemoryManager
from strands_postgres_store import PostgresMemoryStore

store = PostgresMemoryStore(name="user-memories", url="postgresql://user:pass@localhost:5432/db")
agent = Agent(memory_manager=MemoryManager(stores=[store]))

await store.add("The user prefers dark mode", metadata={"kind": "pref"})
hits = await store.search("what theme does the user like?")

How it works

• Semantic recall via pgvector: ORDER BY embedding <=> query (cosine distance) over an HNSW index — native ANN in Postgres.
• You bring the embeddings. Default embedder is Amazon Bedrock Titan Text v2 (1024-dim, cosine); pass any embedder callable for OpenAI / Cohere / local models.
• Each add stores a row {id, content, embedding vector, metadata jsonb, created_at}; search embeds the query and runs the ANN search, surfacing the distance as _score.
• The table, the vector extension, and the HNSW index are created automatically.

Requirements

The pgvector extension must be installed on the server. On Debian/Ubuntu: apt install postgresql-16-pgvector; on macOS (Homebrew): brew install pgvector; managed services (RDS, Cloud SQL, Azure) expose it as an extension. The store runs CREATE EXTENSION IF NOT EXISTS vector on init.

Configuration

PostgresMemoryStore(name, url=None, *, table_name="strands_memory", engine=None, description=None, max_search_results=None, writable=True, extraction=None, embedder=None, dimensions=1024, region_name=None)

The memory store (strands-postgres-store) is distinct from the byte storage backend (strands-postgres-storage). Also published as strands-store-postgres.

License

MIT