isanna bia

What it is

Coding agents forget everything between sessions. Bia is the substrate that fixes that: store, search and govern long-term memory over Postgres with pgvector, exposed over the Model Context Protocol so any MCP-capable client — an editor, a terminal agent, a CI job — can read and write it.

The open substrate is Apache-2.0 and contains no license check of any kind: not at import, not at startup, not at request or write time. It runs on your own infrastructure with no account and no phone-home. The hosted tier adds governance on top — signed erasure receipts, audit retention and export, curation, team policy, SSO — and the substrate underneath never degrades when you decline it.

How it works

Storage

Postgres plus pgvector. Memories carry provenance, not just text, so a recalled claim can be traced to who wrote it and when.

Interface

MCP, so the client integration is the protocol rather than a per-editor plugin.

Isolation

Tenant separation is proven by tests that boot a clean database and attempt to cross the boundary, rather than asserted by convention.

Governance

Audit-log retention with a scheduled purge, an attestation endpoint that reports what the deployment actually enforces, and signed receipts for erasure — so a deletion request produces evidence, not a promise.

What is true today

In service. The substrate is open source and the hosted half runs at bia.isanna.ai. Retrieval quality is the current line of work — the research behind it is on the engineering page.