Agents Retrieve but Do Not Re-Derive: Coherence Costs Concentrate on Conclusion

When agents share a memory store, one agent's action can leave another agent's stored decision stale. We ask when a stale decision still drives action. Across about 18,000 trajectories on three backbones (Qwen 3.5 9B, GLM, Claude Opus 5), agents keep acting on a stored decision unless something invalidates the decision itself. When only the premise changes, Qwen follows the stale decision in 100% of runs. Adding the current value to the record does not help: Qwen and GLM still follow the stale decision in 100% and 89% of runs. Adding a note that the old decision is superseded cuts this to 1% and 0%. Opus 5 never follows a stale decision, and it re-derives the correct one in 98% to 100% of runs once the decision rule is in context. In an 8-entry store with partial staleness, the smaller models stay stale on 33% to 76% of stale entries even with the rules in the prompt. We propose GuardWrite, which attaches validity predicates at write time and rewrites a decision whose predicate fails. Extracted predicates catch every invalidating change whenever the extractor returns usable output. The full protocol gives 100% correct action at every staleness level.