Retrieve for relevance only after checking scope, validity, and revocation

Attach scope, provenance, source revision, observation time, and validity status to records. Before use, apply the relevant access and task restrictions and check current revocation or supersession state. A record can be useful historical evidence while being ineligible as a current instruction or fact. If an index returns stale candidates, a final authoritative filter can reject them and retrieve more. Measure how often that happens, because an apparently full top-k response may shrink to very little usable evidence after the actual validity checks.

A time-to-live is one freshness policy, not proof that a record remains true throughout its lifetime. A source may change or be revoked earlier. Conversely, an old mathematical identity may remain applicable even after a generic recency score decays. Define freshness by record type and the action that will consume it. A derived memory needs its source dependencies reconsidered when they change; removing the original row alone does not remove an unsupported summary elsewhere. Provenance and invalidation systems model these relations explicitly rather than relying on similarity to discover them later.

A system can rank eligible records using relevance, recency, and importance, as the original Generative Agents explores in its simulated environment. Weights and feature normalization determine that heuristic. An embedding similarity near one does not mean the claim has ninety-nine-percent probability of being true, and adding a recency term does not turn the total into a Bayesian belief. Keep score components and their purpose visible. A different query, normalization population, or stored corpus can change rankings without changing the underlying factual evidence.

Retrieval also needs diversity and source dependence checks. Several near-identical passages can crowd out a contradictory update or an exception required by the task. Deduplicate logical sources and inspect the actual context assembled for the model. If a source cannot be resolved, preserve that failure rather than presenting its cached summary as equally current. Immediately before an effect, recheck required revisions or use an atomic precondition where supported. This closes a particular check-to-use gap; it does not promise that all external state remains fixed afterward.

Test a record revoked while it remains in an approximate index, a source changed after retrieval, a conflicting newer record, and an inaccessible original source. Verify that summaries, prompt caches, and action preconditions follow the declared invalidation policy. Measure usable coverage, backfill cost, and downstream errors by record age and scope. A system that only tests fresh successful reads can miss the most consequential failure path. Research retrieval mechanisms should be compared under these controlled conditions rather than assumed correct because memory improved an unrelated benchmark.

The cache lab isolates reuse without any validity metadata. Reset, set both C=4 and s=4, predict four misses and twelve hits at the initial sixteen requests, then inspect or Play. Reset and set both C=2 and s=4 for sixteen misses. The empty fully associative LRU model accesses(i×s) mod 16, adds two requests per second, and caps at 4096. It has no source revision, query scope, TTL, revocation, or corrective update. Its hit count therefore cannot establish that a real retrieved memory remained usable.