AI agents have lifespans, AgingBench measures them โ a longitudinal reliability foundation for agent lifespan engineering.
We're looking for collaborators with production agent traces, sponsors for larger-scale benchmarking, and contributors with new scenarios for agent lifespan engineering.
Long-lived AI agents are increasingly deployed as persistent operational systems, yet they are still evaluated like freshly initialized models. Day-one benchmarks miss a basic systems question: how long does an agent remain reliable after deployment? Even when model weights are frozen, an agent's effective state keeps changing as it compresses interaction history, retrieves from a growing memory store, revises facts after updates, and undergoes routine maintenance. Reliability therefore becomes a lifespan property of the full agent harness, not only a snapshot property of the base model.
We introduce AgingBench, a longitudinal reliability benchmark for agent lifespan engineering: measuring not only whether deployed agents degrade, but what form the degradation takes and where repair should target. AgingBench organizes agent aging into four mechanisms: compression aging, where write-time summarization drops future-relevant details; interference aging, where accumulated similar memories crowd out the target fact; revision aging, where changed or derived state is not updated correctly; and maintenance aging, where lifecycle events such as flushing or recompaction trigger regressions. To diagnose these failures, AgingBench uses temporal dependency graphs and paired counterfactual probes that produce diagnostic profiles for the write, retrieval, and utilization stages of the memory pipeline.
Across 7 scenarios, 14 models, multiple memory policies, and both runner-controlled and autonomous agents, over ~400 runs spanning 8 - 200 sessions show that agent aging is not one-dimensional: behavioral tests can remain clean while factual precision decays; derived-state tracking can collapse sharply within a single model; and the same wrong answer can require different repairs depending on what the diagnostic profile points to. These results suggest that reliable agent deployment requires lifespan evaluation, mechanism-level diagnosis, and stage-targeted repair, not only stronger day-one models.
Three key questions:
AgingBench is a paper, a leaderboard, and a runnable benchmark. Pick the door that matches what you want to do next.
One command, ten minutes. Three release modes (Lite / Full / โถ Lifespan Check) and a plug-and-play surface.
โ Get startedMulti-track results: model swaps, custom memory policies, runtime controllers, autonomous agents.
โ See results For depthSeven scenarios in detail. AgingCard schema. Counterfactual diagnosis. Contributing and roadmap.
โ Read the docsAcross scenarios, models, and memory policies, agents that pass day-one evaluation often show longitudinal degradation across sessions. See more results in our evaluation โ
Decay is rarely a single phenomenon. AgingBench organizes the observed failure patterns into four mechanisms, each pointing to a different way reliability can erode. Read the deep version โ
Write-time abstraction destroys information before future queries are known. Low-frequency details (numbers, names, constraint values) are discarded; high-level summaries survive. The system "remembers what it was about" but loses what it actually said.
Even when no information is lost and nothing has changed, growing state buries the relevant fact behind similar entries during retrieval. Orthogonal to revision: freezing all facts does not prevent it.
The agent fails to track changing truth. Especially severe for dynamic latent state (a budget, a counter, a configuration) where the answer is derived from accumulated updates. A single missed update silently contaminates every subsequent query.
Routine operational events (recompaction, prompt updates, log cleanup, model swaps) silently alter behavior. Driven by actions on the agent, not by its interaction with memory. Causes performance cliffs invisible to standard evaluation.
Frozen weights, growing memory. Each tab walks one agent through a real AgingBench scenario and shows how the same query tends to degrade across sessions, for different reasons. Vignettes are drawn from the actual scenario configs in the benchmark repository.
Knowing how long and how a system tends to decay is not yet enough to act on. The third ALE question asks where in the memory pipeline a failure originates โ the candidate repair site. AgingBench instruments the four pipeline components (๐ฒ write, ๐ฎ store, โ retrieve, ๐ฐ utilize) and uses paired counterfactual probes that swap each component with an oracle to surface a stage-level diagnostic profile.
Four candidate repair sites โ ๐ฒ (write/compression), ๐ฎ (store), โ (retrieve), ๐ฐ (utilize). Counterfactual probes swap each with an oracle to localize where a failure is consistent with.
Aggregate failure rates can look nearly identical across scenarios while the underlying stage-level profiles differ. Two systems with the same accuracy can point to different repair targets.
Selected results, each anchored on a results figure โ scroll the strip to browse. See our paper for detailed findings and intervention implications.
Across the seven scenarios, headline metrics tend to trend downward over the deployment horizon for the (model ร memory policy) combinations we tested. The rate and shape vary by mechanism, and none of the configurations we evaluated avoided the trend.
S1 half-life heatmap: switching the compaction prompt produces a larger half-life delta than any model swap in the row. The memory layer dominates, not the model.
On S2, constraint violation rate stays at zero throughout deployment while constraint precision drops and lag recall collapses. Behavioral monitors miss the decay entirely.
Across 7 models on S2, accumulator error and forget accuracy do not co-improve. Larger models are not consistently better โ revision is representational, not a capacity problem.
On S6, flush / recompact / early-shock variants share the pre-shock window but produce distinct post-shock recovery shapes โ the type of routine event matters more than the model running underneath.
@inproceedings{agingbench2026,
title = {Your Agents Are Aging Too: Agent Lifespan Engineering for Deployed Systems},
author = {Zhu, Jianing and Ro, Yeonju and Robertson, John and Wang, Kevin and
Li, Junbo and Vikalo, Haris and Akella, Aditya and Wang, Zhangyang},
booktitle = {arXiv preprint arXiv:2605.26302},
year = {2026}
}