Known limits

Meaning is still the agent’s judgment

Rationale proves which Records govern a target and reports lexical intent conflicts with a conservative polarity hint. It does not call a remote model to decide meaning. The hint is noisy in both directions — it can flag an unrelated rule as opposed and miss a direct contradiction — so every conflict is marked as unverified overlap, and undetermined never becomes a blocking verdict. The governing Records themselves are reliable; read those.

Structure is as good as the provider

Codebase Memory is optional and fallible. Some languages resolve calls by name, so an observed edge means “present in the index”, not a semantic guarantee. Relationship states are derived on every call and never delete an explanation: orphaned and unknown ask for a look, not a cleanup.

Capture is autonomous

Agents write Records without an approval queue. The gate removes noise and duplicates, and provenance always says an agent asserted it, but the quality of memory still depends on the agent following the protocol. Pin the rules that must not move, and review .rationale/records/ in pull requests like code.

Skill selection is probabilistic

An agent decides whether to load the rationale skill from its description, and it can miss. The protocol in CLAUDE.md or AGENTS.md is the floor every conversation keeps; /rationale and $rationale make the skill explicit. The skill’s evals measure selection and behavior, but they are run by maintainers, not on your machine.

One repository at a time

A project is one Git repository. There is no multi-repository federation, no hosted canon, and no account.

Governance documents in progress

Some architecture decisions behind 1.0 (record lifecycle, local data exclusion, user-scoped registration, the activity stream) are recorded as ADRs whose status is still proposed pending independent review. The behavior is implemented and tested; the formal acceptance is open.