Cambium sets governance rules for agent-written memory

Cambium proposes a governance standard for the corpora that AI agents write to over time, per its GitHub repository. Once an agent's memory store becomes a product's source of truth, nobody's tracking where each fact came from or whether it's still accurate. Nobody's tracking who approved the change that put it there, either. Cambium's schema covers exactly that: provenance, review, and drift.

The project, from developer KimGLee, sits at 204 stars after a push on August 7.

It's not alone. bybit-exchange/kaas, an MIT-licensed knowledge-base compiler with 88 stars, turns notes and transcripts into a queryable Markdown wiki exposed over MCP. It skips embeddings entirely, per its repo.

That's the detail worth sitting with. Vector databases became the default way to give an LLM memory because they're fast to query. A float in a vector doesn't tell you why the agent believes something, the way a provenance record does.

kaas gives up embeddings for a wiki a person can actually read. Cambium gives up an unstructured blob for a schema built around who wrote what and who reviewed it.

Neither project says how it handles conflicting edits when two agents write to the same document at once. And neither has the adoption yet to prove the pattern generalizes past its own repo.

Skipping embeddings on purpose is the tell. Vector stores win on retrieval speed and lose on the thing agent memory needs most once it's editing itself: an audit trail. Watch whether more agent-memory tools drop embeddings the way kaas already has, trading speed for something a human can actually read.