Memory needs
governance.

Once an autonomous system can carry context forward, a second problem appears. What exactly has been remembered? Who or what does it refer to? Which source supplied it? Is that source authoritative for this fact? Is the context still current? Who may see it, for what purpose, and who remains responsible for deciding what happens next?

Storage answers none of those questions. A database can preserve a claim long after its usefulness has expired. A model can compress contradictory observations into a confident summary. An application can copy information into a workflow without carrying forward the conditions under which it was collected.

Governed memory keeps the qualifying context attached. Identity, source, time, confidence, policy, purpose, correction and human review are part of the memory rather than optional metadata around it.

A memory becomes dependable when the system can show what it knows, why it may rely on it and where human authority begins.

Physical environments contain overlapping identifiers. A resident may be represented by a room label, a local application record, a wearable, a voice profile, a family reference or a temporary hospital identifier. Devices and robots maintain their own local identities. Places change purpose. People move.

A resident graph is intended to connect these references without pretending that every match is certain. Resolution should preserve candidates, confidence, source evidence, corrections and the distinction between an account, an actor, a person and the subject of context.

When ambiguity remains material, the safe result is not a convenient guess. The system should constrain disclosure and action, request clarification or route the case for human review.

Governed memory needs more than a creation timestamp. It should preserve when something was observed, when it was received, the period for which it is considered valid, the policy that controls decay and the event that should trigger revalidation.

Different context decays at different speeds. A stable preference may remain useful for months. A location may become stale in seconds. Capacity, consent, risk, medication instructions and delegated authority may change according to their own rules and circumstances.

The point is not to delete history when current validity ends. Historical context can remain valuable evidence. The point is to stop old context quietly presenting itself as current authority.

Operational truth is often contested. A resident's stated preference may differ from an older record. Two sensors may observe different conditions. A model may infer a change that a human has not confirmed. A corrected identity may invalidate context previously attached to it.

Flattening these states into a single current value destroys the information needed to govern the next step. A context layer should retain competing claims, their sources, timing, confidence, truth lane and correction lineage. Policy can then decide whether one claim is authoritative, whether the context must be withheld or whether human review is required.

Uncertainty is not an implementation failure. Hidden uncertainty is.

Dumping an entire resident history into an agent, robot or interface is not continuity. It increases disclosure, confusion and attack surface while making it harder to see which facts actually governed the outcome.

A context slice should be compiled for the actor and task. It can include relevant identity, current facts, applicable preferences, constraints, source and freshness indicators, policy outcome, confidence and the required human review path. Everything else remains behind its proper boundary.

Typed contracts, deterministic tests and architecture controls are meaningful evidence of design progress. They are not proof of live operation, safety performance, certification or regulatory status. Read current product statements alongside the canonical facts and proof boundary.