#66 PostgresStore: share one memory across processes and machines (1.0, item A)

[Anonymous]

Originally created by: Thanatos9404

Why

SQLite is one file on one machine. Production agents run several workers, serverless functions or replicas, and teams already run Postgres (Supabase, Neon, RDS). Without a shared store, llmslim serve can't scale past one box, and a hosted tier can't exist.

What

pip install "llmslim[postgres]" (psycopg 3). Then:

Memory("postgresql://user:pass@host:5432/db")              # or PostgresStore(url, table="my_app_memory")

:::bash
llmslim serve --db postgresql://user:pass@host:5432/db     # same for llmslim mcp

• Schema: the table (llmslim_memory by default, or table=...) and <table>_meta are created if missing.
• Creation runs under pg_advisory_xact_lock, so several processes can start at once (tested with 6 at once).
• A newer schema version is refused rather than misread.
• Inputs: a URL, for which the store opens and later closes its own pool; your own psycopg_pool.ConnectionPool, which it leaves open; or an autocommit psycopg.Connection. A non-autocommit connection is refused.
• Times are TIMESTAMPTZ and come back as UTC-aware datetimes even when the session TimeZone is not UTC. Tested with Asia/Kolkata and microseconds kept.
• Order: seq is an identity column. Eight threads × 25 writes kept all 200 turns with unique, ordered seq, and each writer's turns stayed in the order it added them.
• Duplicate ids raise ValueError, and the whole batch rolls back.
• Safety: table names are validated before connecting (no SQL injection through table=), and all SQL is composed with psycopg.sql.Identifier.
• --db: in serve and mcp it accepts a postgresql:// URL. A shared storage_target() helper keeps SQLite paths working as before.

Tests

• The whole tests/test_memory.py suite now also runs on Postgres when LLMSLIM_TEST_POSTGRES_URL is set: 16 parametrized tests across the 3 stores.
• tests/test_memory_postgres.py adds 13 more:
• URL detection;
• table-name validation;
• persistence;
• duplicates and a newer schema;
• pool and connection inputs;
• time zones;
• concurrent writers;
• concurrent schema creation;
• the HTTP server on Postgres.
• New CI job, "Memory on Postgres": it runs both files against a postgres:17 service.
• Locally, against PostgreSQL 18:
• 1,801 passed, 8 skipped;
• postgres.py coverage 98%;
• every test table was dropped afterwards.
• Without a database, the Postgres-only tests skip and the rest run.

[Anonymous]

Originally posted by: vercel[bot]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
llmslim	Ready	Preview	<relative-time datetime="2026-10-17T09:41:29.438Z">Oct 17, 2026 9:41am UTC</relative-time>

[Anonymous]

Ticket changed by: Thanatos9404

• status: open --> closed