Agent memory eviction causes many irreversible errors in AI question answering
What Eviction Destroys: A Restore-Counterfactual Audit of Forgetting in Agent Memory
Computation and LanguageArtificial IntelligenceDatabases
Summary
AI systems that remember past information sometimes have to forget parts of it because they can only keep so much. The authors studied exactly what kind of mistakes happen when AI has to remove old information from memory. They found that most errors caused by forgetting can't be fixed later, especially when the memory is very limited. This shows how important the way AI decides what to forget is for keeping good answers. Their method lets us see which mistakes happen because something was truly lost versus just temporarily hard to find.
agent memoryinformation evictiontoken budgetretrievalrestore counterfactualFIFO policyLLM-importance evictionoracle-answerable errorGPT-4o-miniconversational AI benchmarks
Authors
Chen Shen
Abstract
Agent memory systems must discard stored information when their history exceeds a fixed token budget. Existing budget-accuracy frontiers quantify the resulting loss in accuracy, but do not distinguish irreversible losses caused by eviction from recoverable retrieval failures. We introduce the restore counterfactual, a per-question paired intervention that reinstates the question's gold evidence in the read-time context and reruns the same reader. Combining the change in correctness with whether the evidence was retained after eviction classifies each oracle-answerable error as recoverable, irreversible, or residual; in the residual case, the answer remains incorrect after restoration. We evaluate FIFO, random, redundancy-aware, and LLM-importance eviction on LongMemEval-S at three budgets and under two retrieval regimes, using GPT-4o-mini as the primary reader and judge and GPT-5.4-mini as a robustness reader. Under top-k retrieval at an 80k-token budget, the irreversible share among errors corrected by restoration is 0.67-0.73 for FIFO, random, and redundancy-aware eviction, compared with 0.60 for LLM-importance. At 8k tokens, it reaches 1.00 for all four policies. Recoverable errors occur under top-k retrieval at 80k tokens but are absent under forced-gold injection by construction, so budget-accuracy results are not directly comparable unless the retrieval regime is reported. An exploratory matched-accuracy analysis detects no difference in irreversible rate among accuracy-matched policy pairs at a resolution of 1.2-6 percentage points. The same analysis detects the deliberately destructive control. To our knowledge, this is the first per-item, per-question restore-counterfactual audit of eviction for external agent-memory stores on a standard conversational benchmark.