Provenance-Tiered Conflict Resolution in Agent Memory

We use the term agent context integrity for the property that the information an autonomous agent communicates, stores, retrieves, and acts upon preserves enough provenance, attribution, verification, and conflict semantics to support reliable reasoning, and study one component of it: conflict resolution in long-lived memory. Autonomous agents increasingly answer from an accumulated store of facts rather than from a single retrieved passage, and that store, like any long-lived memory, comes to hold facts that conflict: a value the agent recorded last week, a newer guess it made this morning, a decision a human once handed it. When asked a question whose answer is contested, the agent must resolve the conflict, and the store gives it only the information it kept. We observe that mainstream agent-memory systems keep recency and relevance but discard the one signal that determines which conflicting fact to trust: where the fact came from. We propose attaching a discrete provenance tier to every fact at write time, human-decided > source-verified > agent-asserted, and resolving conflicts by tier rather than by recency. We give two closed-form models for the accuracy of a recency-resolving store versus a tier-resolving store as a function of conflict depth and conflict density, confirm them with an exact model-free simulation, and measure the effect on five instruction-tuned models across three architectures. The deterministic model is a scaffold, not the finding: under a deliberately adversarial construction it fixes the extremes (a recency-only store cannot resolve an adversarially-ordered conflict; a tier store resolves it by construction), setting up the question the empirical study answers, whether real models exhibit the same pull toward recency when the tier is withheld. That empirical result is directional across every model tested, though its magnitude depends on how well a given model engages the task, and it is where the paper’s weight rests. We situate this against trust-weighted retrieval, where reliability-based conflict resolution already exists for external web sources, and argue the contribution is its move into agent memory as a discrete, write-time, unforgeable tier: in particular the highest tier is minted only through an authenticated owner channel, so an agent cannot promote its own assertion to the tier that wins. We frame the broader property, verified at the source, attributed to its origin, resolvable to ground truth, as context integrity, and position this as one measurement of it.