acceptodds
Desk rejected at ICLR 2027

MIRAGE: Memory Revision and Path-Aware Retrieval for Conflict Resolution in LLM Agent Memory

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

LLM agent memory systems increasingly rely on non-parametric state to reuse experience across tasks. However, memory maintenance remains under-specified: append-only storage can leave stale evidence, conflicting facts, and disconnected retrieval snippets in the active context. We study memory-state maintenance as an alternative to append-only context construction: deciding what evidence should remain active, what should be suppressed, and what structure is needed for multi-hop reuse when memory updates follow an explicit authority signal. We propose MIRAGE, a memory-maintenance framework that maintains agent memory through ordered revision, provenance marking, and path-aware retrieval. In this paper, we evaluate three concrete MIRAGE paths on MemoryAgentBench FactConsolidation: ordered conflict revision (MIRAGE-R), provenance-marked revision (MIRAGE-P), and graph/path-aware retrieval (MIRAGE-G). On the 6k multi-hop setting, MIRAGE-R improves accuracy on 100 questions from 0.28 to 0.68 while reducing token use by 32.8%; on the 32k multi-hop setting, accuracy improves from 0.21 to 0.65 with a 34.5% token reduction. On the same 6k multi-hop setting, MIRAGE-P maintains comparable accuracy at 0.71 while reducing empty outputs, and MIRAGE-G raises retrieval-based accuracy from 0.17 for lexical top- retrieval to 0.59 while preserving an 80.1% token reduction. Our findings suggest that MIRAGE provides a practical memory-maintenance layer for long-context agents, especially in settings where persistent state must be revised, audited, and retrieved through multi-hop evidence rather than simply accumulated.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, memory revision, path-aware retrieval, conflict resolution, agent memory

Related papers

Discussion (0)

Sign in to comment.