Usage and Activation

A memory appearing in search results does not prove an agent used it. A memory delivered to an agent does not prove it affected the answer. And frequent use does not make a claim more accurate. The maintained-memory plan treats usage, retrieval priority, and belief as separate signals.

Examined, delivered, and used

The plan distinguishes at least three events:

• Examined: a candidate participated in a retrieval process. This can be counted in bounded aggregate form; it does not require writing a durable row for every candidate on every search.
• Delivered: content was actually returned to a caller through a user- facing get or recall result. A metadata or status lookup is not content delivery.
• Used: an agent or caller explicitly reports that a memory informed its work. This is an optional signal and is not inferred merely because the memory was returned.

Events are attributed to the exact memory version that was involved. That matters when a memory is later refined, merged, or split: old usage should not be copied wholesale to every resulting version. Optional use reports are planned to be idempotent and rate-bounded so retries do not inflate counts.

Usage records have their own retention and privacy rules. The plan proposes a 30-day default for detailed usage events, with bounded canonical checkpoints to rebuild longer-lived summaries. Once detailed events expire, summaries can still describe aggregate use, but they cannot recreate an exact event history. Erasure must remove associated usage data as well.

Usage does not establish support

Usage is not evidence that the content is true. Reading a claim, delivering it many times, or reporting that it was useful cannot promote a hypothesis to a supported or corroborated belief. Confirmation, correction, or corroboration must arrive as evidence contributions with their own provenance and acceptance rules.

This prevents a popularity loop in which a claim becomes more prominent, is therefore returned more often, and is then mistaken for a more reliable claim because it was returned often. Belief labels continue to depend on evidence, not on retrieval counts.

Activation may break a relevance tie

The planned activation prior estimates how readily a memory may come to mind from bounded usage summaries and task context. Its initial role is deliberately narrow: when two results have equal fused retrieval scores, activation may order that tie. It should not change the underlying retrieval scores, reorder unequal relevance scores, or override deterministic final tie-breaking.

This is an accessibility signal, not a truth signal. A well-used memory can still be false, historical, unsupported, archived, or hidden by privacy rules. Activation cannot restore it to current results or bypass a revision barrier. Likewise, an old but highly relevant fact should remain retrievable even if it has not been used recently; the plan calls for measuring cold relevant facts before enabling the prior.

The initial decay rule is described as ACT-R-inspired, with a provisional operator-configurable parameter. It is not presented as a value fitted to Remem users. The exact calculation and defaults are left for implementation and evaluation, and any future change to ranking policy needs its own review.

For how results carry status and uncertainty, see Belief-aware Recall. Usage can affect a narrow ranking tie-break there; it cannot alter the belief labels described in Beliefs and Evidence.