dataqbs

When Should LLM Agents Forget? Memory Filtering Under Query-Critical Resolver Errors

· Source: arXiv cs.AI

Large language model (LLM) memory systems routinely purge data deemed outdated. This study shows that overly strict deletion can hurt answer accuracy when essential facts are removed. The authors evaluated three distinct models and applied a controlled sensitivity intervention: moving from perfect retention to the lowest tested reliability, the benefit of hard deletion shifted from gains of +0.186 to +0.230 to losses of –0.176 to –0.181. Targeted experiments revealed that false‑drop errors degrade performance far more than false‑keep errors, though a never‑delete strategy can also fail if incorrect labels influence text generation. In a frozen calibration‑test split (102/102), a deployment‑aligned decision gate achieved an accuracy of 0.655, slightly better than always‑delete (0.641) and close to an oracle‑guided selection based on test data (0.662). Global precision cross‑overs depend on reference labels and vary by model and task, so they cannot serve as deployment thresholds. The authors advise treating deletion as a cost‑sensitive decision: validate the risk of removing critical facts in data resembling the target environment, or otherwise retain information when such validation is unavailable. This issue is important because it directly impacts the reliability of LLM‑based applications that rely on accurate memory, such as virtual assistants or recommendation systems.

Read the original article on arXiv cs.AI

This summary is an informational synthesis produced by dataqbs.com. All rights to the original content belong to its author and the cited media outlet. We act solely as curators of technology news and claim no authorship.

Read this in Español · Deutsch · Français · Italiano