---
format: "aidr-story-markdown/v1"
id: "f5dc4cd120b45659518fb591bab6aa27c9c58619caff9fd1deb945dff07a13a6"
canonical_url: "https://aidr.today/f5dc4cd1?lang=en"
title: "When Should LLM Agents Forget? Memory Filtering Under Query-Critical Resolver Errors"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-16T05:00:00.000Z"
category: "Research"
topics: ["llm","anthropic"]
source_urls: ["https://arxiv.org/abs/2611.17338"]
summary: "Hard-filtering LLM memory can reduce downstream accuracy when resolver errors delete query-critical facts. A simple deployment-matched gate reached 0.655 accuracy on a frozen 102/102 calibration/test split versus 0.641 for always hard-drop. Deletion should be treated as cost-sensitive by validating query-critical false-drop risk on matched data or defaulting to label-free retention."
---

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

> [Open the canonical story](<https://aidr.today/f5dc4cd1?lang=en>)

**Published:** 2026-11-16T05:00:00.000Z
**Category:** Research
**Topics:** llm, anthropic

## Summary

Hard\-filtering LLM memory can reduce downstream accuracy when resolver errors delete query\-critical facts\. A simple deployment\-matched gate reached 0\.655 accuracy on a frozen 102/102 calibration/test split versus 0\.641 for always hard\-drop\. Deletion should be treated as cost\-sensitive by validating query\-critical false\-drop risk on matched data or defaulting to label\-free retention\.

## Sources

- [Story source](<https://arxiv.org/abs/2611.17338>)

