Jaybase

Most agent systems treat the data store as an afterthought — a Postgres table or a key-value cache that gets mutated silently and leaves nothing auditable. Jaybase is purpose-built for the opposite constraint: every fact an agent writes is appended as an immutable record, never overwritten. The core workflow is write-once, read-sequentially: the agent appends a fact, that fact is timestamped and stored, and any downstream process — audit, replay, approval gate — reads the ordered log. No deletes. No silent updates.

The differentiating feature is replayability tied to non-determinism. When agents produce different outputs on different runs — different prices, different approval decisions, different classifications — a mutable store makes those divergences invisible. Jaybase makes them explicit: every fact is a permanent record, so you can walk the exact sequence of what the agent decided, in order, and verify or dispute it. The docs describe this as the primary design target for regulated environments.

Jaybase fits single-tenant business process automation where the agent has write access to something that matters: accounting entries, compliance records, approval chains. It is an open-source library under AGPL, self-hostable, with an API surface. It does not fit multi-tenant SaaS architectures without significant isolation work, and teams needing rich query patterns — aggregations, filtering, joins across fact types — will find the append-only model a friction point rather than a feature. The project carries a small community footprint, so teams should expect to write their own integration layers for anything not covered in the README.