← News

Agents & MCPSources: arXiv · 16 Nov 2026

Study finds hard-filtering stale memories can hurt LLM agent accuracy

An arXiv abstract reports that deleting facts labelled stale can backfire when resolver errors remove query-critical information, and proposes cost-sensitive deletion instead.

In plain words

Some systems for large language model agents automatically delete memories they classify as outdated. The paper's authors show that when this deletion step wrongly removes a fact the current query needs, accuracy can drop rather than improve. They suggest treating deletion as a risk decision: validate it on matched data, or keep memories by default.

What was announced

The abstract reports a controlled sensitivity intervention across three models: hard-drop moved from gains of +0.186 to +0.230 at perfect resolution to losses of -0.176 to -0.181 at the lowest tested reliability. Paired targeted interventions indicated that false drops caused significantly greater degradation than false keeps, though a non-deleting baseline could still fail when incorrect labels steered generation.

On a frozen 102/102 calibration/test split, a simple deployment-matched gate reached 0.655 accuracy versus 0.641 for always hard-drop and 0.662 for test-informed oracle selection. The authors add that global-accuracy crossovers need gold labels and are model- and task-specific diagnostics, not deployment thresholds, recommending label-free retention when validation data is unavailable.

What it does not show

  • Evidence is limited to the paper's abstract; full methods, tasks and results are not available for verification.
  • The reported numbers come from the authors' own experiments and have not been independently confirmed.
  • The abstract does not name the three models, the memory systems or the tasks, so the findings' breadth is unclear.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.