Consolidation

Consolidation is a planned process for forming a more general memory from related, supported memories. It could help an agent retain a useful pattern across many episodes instead of retrieving every example separately. The generalization must remain connected to the examples and rules that justify it, and it must not quietly replace those examples.

Candidate selection is not proof

The plan starts with memories about the same resolved subject and compatible scope, perspective, and applicability. Vector similarity and graph neighbours may nominate a bounded set of candidate memories. They do not prove that the memories share a rule or can safely be generalized.

Relevant conflicts and exceptions are included as separately labelled context, not counted as support for a generalization. A useful consolidation might propose “usually chooses trains for short trips” from several supported travel decisions, but the examples alone do not establish that rule for every trip, country, or circumstance.

Broader generalizations across subjects need explicit rules and evaluation. Missing evidence is not evidence that no exception exists, so the system must not assume a closed world in which unobserved exceptions can be ignored.

Generalizations begin as hypotheses

Even if every member memory is supported, the inferred common rule is a new claim. The planned acceptance boundary therefore keeps a model-proposed generalization as a hypothesis unless independent evidence confirms it or a registered eligible rule derives it. The model does not gain authority merely because its input examples were individually supported.

The new memory would record its exact premises. If the generalization requires all of them, losing support for one required premise can invalidate that justification. If there are separate independent justifications, each is tracked on its own so one withdrawn path does not erase the others.

Exceptions stay scoped

An exception is not a footnote to discard. A supported exception can defeat a generalization within its declared scope; a contested exception makes that scope uncertain. For example, a general preference may apply except when a particular accessibility need is present. The exception should limit the generalization where that need applies without turning into a universal contradiction.

An exception cannot be derived from the generalization that it defeats, since that would make the conclusion support its own exception. Nor should the system count a fixed number of counterexamples and apply an exception everywhere: scope and evidence determine where it matters.

Members remain available

Consolidation is not destructive deduplication. The original memories and retained episodes remain addressable under their own current and historical rules. A future retrieval prior might lower the prominence of a member after a consolidation, but it cannot delete it or make its evidence disappear. See Usage and Activation for the separate planned ranking signal.

Bounded and evaluated work

The plan proposes both threshold-based and periodic triggers. New eligible memories can mark a subject and context for later consolidation; a count threshold can make it due sooner, while a periodic pass prevents quiet areas from being ignored. Work is bounded, coalesced, and subject to configured provider budgets. If a budget is exhausted, the work remains pending instead of falling back to a weaker acceptance rule.

Evaluation must test that unsupported generalizations remain hypotheses, required premises are enforced, independent justifications survive separately, and scoped exceptions create the right uncertainty or defeat. Similarity alone cannot demonstrate that a generalization is correct.