// the find
smaramwbc/statewave
Open-source memory runtime for AI agents — reproducible, provenance-tagged context bundles instead of query-time retrieval. Apache-2.0, self-hosted on Postgres + pgvector, Python + TypeScript SDKs.
Statewave is a self-hosted memory runtime for AI agents. It ingests raw events, compiles them once per subject into typed memories with provenance, and assembles token-bounded context bundles on Postgres with pgvector. It is aimed at teams running support agents or coding assistants who want the same query to return the same context.
- Compile-once, then assemble is a concrete design choice. Compilation is idempotent, bundles are content-hashed, and receipts are immutable and ULID-addressable, so determinism can be tested rather than just asserted.
- The heuristic compiler is regex-based and runs fully local with no API key, so the whole pipeline can be exercised in CI or on a laptop before any LLM spend.
- Provenance is built in. Every memory points to its source episodes, and receipts carry HMAC-SHA256 signatures and a frozen policy snapshot, which most memory libraries skip.
- Test coverage is broad for a young project, including integration tests for idempotent ingest, tenant isolation, residency pinning, and receipt signing, and the README lists its known limitations plainly.
- The default boot is demo mode with hash-based stub embeddings and the heuristic compiler, so semantic search does not work until you configure LiteLLM. The quickstart does not make that obvious.
- Multi-tenancy is enforced in the application layer with no Postgres RLS, and rate limiting is keyed by IP only. That is acceptable for an internal deployment but needs review before one database serves several customers.
- The case for the approach rests mainly on the project's own benchmarks, and the FAQ concedes that bundles cost more tokens per answer than a plain fact store. Anyone choosing it should measure that tradeoff on their own queries.
- Much of the substance sits in other repos: the docs, SDKs, connectors, and examples each live elsewhere. Evaluating the system means reading across six repositories, and the README reads more like a product page than a technical overview.