Agent memory should consolidate by evidence, not recurrence

Agent memory should consolidate an observation because its evidence became stronger, not merely because the same sentence appeared again. Recurrence is a useful signal of salience, but it is also how duplication, shared-source errors, stale beliefs, and repeated attacks acquire false authority. Counting mentions turns an echo into a fact.

Consolidation is necessary. Raw episodic records accumulate until retrieval quality falls, contradictions coexist without priority, and the context budget fills with near-duplicates. Systems therefore reinforce recalled memories, decay unused ones, distil episodes into semantic facts, and promote recurring action sequences into procedural skills. Those moves separate learning from hoarding. The mistake is treating frequency as the promotion gate.

Five identical observations may be one observation copied through five channels. A CRM export, a support dashboard, and a weekly report might all derive from the same underlying record. Ten agent runs may repeat a preference because each read the same poisoned memory. A user may correct the belief once with direct authority. Frequency voting would preserve the error; provenance-aware evidence should replace it. Memory governance begins by asking who asserted a claim, from which source, under what scope, and with what right to correct or delete it.

I want consolidation to operate on claims with evidence sets. Each claim carries source identity, observation time, valid time, confidence, scope, and derivation lineage. Independent corroboration raises support; copied derivatives do not. Direct user statements can outrank model inference within the user’s scope. Authoritative systems of record outrank summaries. Contradictions remain attached rather than averaged away, and the consolidated view records why one version currently wins. A memory read is a query plan over those evidence relationships, not nearest-neighbour retrieval over flattened prose.

Procedural memory needs an even stricter gate. Repeating a sequence does not prove the sequence is safe, general, or causal. A promotion candidate should survive tests across representative tasks, record preconditions and expiry, and distinguish required steps from incidental choreography. One verified recovery procedure is worth more than twenty successful traces that never encountered failure. The durable skill should link back to its validating episodes so a future incident can challenge it.

Forgetting then becomes evidence maintenance rather than arbitrary decay. Low-value episodes can expire after their claims and provenance have been preserved. Stale claims lose current validity without erasing the historical fact that they were once true. Contradictory evidence can demote a memory immediately instead of waiting for its retrieval count to fall. Staleness is a lifecycle property, so time and source currency must participate in consolidation.

This design makes consolidation testable. Seed the store with independent support, duplicated derivatives, a later correction, and an expired source; then assert the winning claim, retained dissent, and explanation. Track false promotion, false forgetting, and time-to-correction rather than celebrating compression ratio alone. A shorter memory is not necessarily a wiser one. The quality measure is whether evidence changes belief in the intended direction.

There is one precise concession: recurrence can be sufficient for low-stakes personalization when observations are independent, reversible, and cheap to correct—for example, repeatedly choosing compact table layouts. Even there, the system should express the result as a tentative preference, not an identity fact. The promotion threshold should rise with consequence, persistence, and sharing scope.

Designing memory is data modeling because consolidation decides grain, keys, lineage, precedence, and lifecycle. A memory system learns responsibly when it can explain not only what it believes, but why the available evidence justifies believing it now. Recurrence helps find candidates. Evidence earns durability.