pydantic-ai-postgres-memory

A PostgreSQL MemoryStore for the PydanticAI harness Memory capability, built from a connection URL. The harness already ships a transactional PostgresMemoryStore that takes a caller-owned asyncpg pool; this package wraps it so it drops into Memory(store=...) like the other pydantic-ai-*-memory backends, creating and owning the pool for you.

pip install pydantic-ai-postgres-memory

from pydantic_ai import Agent
from pydantic_ai_harness.memory import Memory
from pydantic_ai_postgres_memory import PostgresMemoryStoreFromUrl

store = PostgresMemoryStoreFromUrl("postgresql://user:pass@host:5432/db", table="agent_memory")
agent = Agent("anthropic:claude-sonnet-5", deps_type=Deps,
              capabilities=[Memory(store=store, namespace=lambda ctx: ctx.deps.user_id)])

How it works

• Delegates every call to the upstream pydantic_ai_harness.memory.PostgresMemoryStore, so the compare-and-swap and operation-receipt semantics are exactly the reference ones: each write or delete runs in one transaction with a version sequence and an {table}_operations receipt table. Schema is created on first use.
• The asyncpg pool is created lazily on the running event loop and kept per loop, because asyncpg pools are loop-bound and agents are frequently driven from several loops (run_sync per call, workers, tests). Pass pool= to reuse a pool you own instead; aclose() closes the owned pool for the current loop.
• search is the harness's bounded lexical search over the scope's files.

Docs: https://skamalj.github.io/agentstate-reducer/ · part of pydantic-ai-memory

License

MIT