Project
Bento
Agentic memory in Postgres — the elephant remembers so the goldfish can forget.
- Open Source
- Python
- postgres
- memory
- ai
- rag
- Status
- alpha
Bento is project #001 in Goldfish Labs — agent-grade memory primitives that run on Postgres, not a vector store bolted to an LLM. The thesis: durable, queryable, governable memory belongs in the system of record, not in a separate index that drifts.
Why Postgres
A vector store is an index, not a database. It can find similar things, but it can’t transact, join, version, or enforce row-level security on what it returned. Agent memory is none of those problems — it’s all of them at once. You need the recall of a vector index, the truth of a relational store, the ergonomics of SQL, and the operational maturity of something a bank already trusts. That’s Postgres + pgvector + a small kit of opinionated tables.
What ships
/v1/ask— a single HTTP endpoint that takes a question and returns a grounded answer with citations.- Memory primitives: episodic (events), semantic (facts), procedural (preferences) — three tables, three retrieval modes.
- Row-level security from day one. Multi-tenant by default. No “but in production we’ll add auth.”
- No-spin benchmarks: where Bento is faster, we show it; where it trades footprint for accuracy, we say so.
White paper — Bento Under Test
A controlled evaluation of Bento’s Postgres-native memory and retrieval layer: where it wins, where it trades footprint for accuracy, and exactly how every number was produced. All figures come from executable graders (not an LLM judge), reproducible from committed seed records and reported with 95% confidence intervals and a threats-to-validity section.
What’s on this page
Every piece of Goldfish content tagged project: bento shows up below — the blog posts that explain the thesis, the tutorials that get you to a working memory store, the videos and podcasts where we talk it through, and the sample code you can paste into a project today.