Age Is Not Evidence of Staleness: A Recency Decay Is a One-Rate Survival Model, the Rate Belongs to the Fact Rather Than the Clock, and the Silent Change Only a Rate Could Catch Is the Case No Agent-Memory Benchmark Scores

Abstract

Language agents that keep memories across sessions usually discount a record by how long ago it was written or last read. That discount was introduced to decide what an agent attends to, but it is increasingly read as a statement about whether a record is still true. This paper examines that second reading. Taken as a validity claim, an exponential recency term is a survival model with a constant hazard and a single rate shared by every record in the store. The time-sensitive knowledge literature, together with more than two decades of web-crawl measurement, contradicts the single rate: how fast an answer goes out of date differs by orders of magnitude across kinds of fact, and on the benchmark that splits questions by that rate, accuracy falls sharply with it. The paper then complicates the obvious remedy of attaching the rate to the predicate. Two published operationalisations of volatility, one by relation type and one by per-fact edit history, reach opposite conclusions on the same question, and within-class variation is large. It shows by standard derivation that pooling heterogeneous rates makes the pooled hazard fall with age, so a long-unchanged value is evidence of stability. Age since last confirmation and duration held before it therefore point in opposite directions, and a single timestamp cannot separate them. Finally, it observes that the systems and conversational-memory benchmarks located here invalidate a memory only when a later observation contradicts it. The case where only a volatility prior could help, a change the agent never hears about, is scored by no conversational-memory benchmark located here, and the one controlled study that withholds a change and budgets verification does not vary elapsed time or how volatile the fact is. It closes by stating a decision structure as an algorithm and specifying a benchmark that would settle the open parts.

About the author

Pranay Mahendrakar is an ai specialist and large language model engineer based in Bengaluru, India. He builds production artificial intelligence systems and publishes open-access research on how those systems fail. See all 74 papers by Pranay Mahendrakar, or his ORCID record.