MEMOIR Lifts LLM Agent Recall 11.9pp on LoCoMo QA Benchmark

A November 5, 2026 arXiv paper introduces MEMOIR, a training-free memory lifecycle for long-horizon LLM agents that raises evidence recall by 11.92 percentage points over cosine-similarity baselines on the LoCoMo QA benchmark. The system, developed by researchers Mayank Anand, Subham Raj, Aman Vaibhav Jha, and Sriparna Saha, addresses a fundamental failure mode in agent memory: consolidation can silently destroy evidence that future queries require, while retrieval continues to favor stale memories.

Officially released on November 5, 2026, the paper tested MEMOIR on 1,905 deduplicated LoCoMo QA pairs and recorded zero losses across 100 independently audited memory merges.

How MEMOIR Works

MEMOIR combines two core mechanisms: judge-gated consolidation that screens for evidence preservation before any merge, and retrospective attribution that updates retrieval priority after each query. A UCB-style exploration bonus keeps under-exposed memories in play, preventing the system from over-relying on frequently accessed but potentially stale information.

The paper states the design principle bluntly: attribute usefulness after use, and screen for preservation before compression. Existing systems write memories prospectively with static importance scores that never get revised, creating coupled failure modes where consolidation destroys needed evidence while retrieval favors outdated memories.

Benchmark Performance Gains

On 1,905 deduplicated LoCoMo QA pairs, MEMOIR raised evidence recall at rank 5 by 11.92 percentage points over bare cosine similarity. Confidence-gated feedback, the component that runs without reference answers or reranker training, added another 1.50 points on top of base retrieval performance.

An oracle variant flagged as deployment-unavailable added 1.84 points overall and 8.05 points specifically on multi-hop queries. Judge-gated merging shifted text-evidence recall from 58.82 percent to 59.08 percent while shrinking the active memory store by 1.42 percent, with zero observed losses in 100 audited merges and a one-sided 95 percent upper bound on loss rate of 2.95 percent.

Memory Consolidation Without Loss

The judge-gated screening mechanism tests whether answer-relevant evidence is preserved before executing any memory merge. This pre-merge validation ensures consolidation does not silently discard information that future queries will need. The system logged zero losses across 100 independently audited merges, demonstrating that screening gates can prevent the silent evidence destruction that plagues existing agent memory systems.

Active memory storage shrank by 1.42 percent across these merges, showing that MEMOIR compresses the memory footprint while maintaining retrieval quality. The retrospective attribution mechanism updates retrieval priority based on which memories actually supported answers, rather than relying on static importance scores assigned at write time.

Training-Free Deployment Context

MEMOIR requires no retraining or fine-tuning of the underlying language model, making it a drop-in addition to existing long-horizon agent architectures. The confidence-gated feedback component operates without reference answers, allowing the system to learn from production queries in real time.

The paper arrives amid a wave of memory-and-RAG agent research where retrieval failures, rather than model size, increasingly bound long-horizon agent performance. MEMOIR addresses this bottleneck by coupling consolidation screening with post-query attribution, ensuring that compression decisions account for actual usefulness rather than prospective guesses.

What This Means

MEMOIR demonstrates that training-free memory lifecycles can deliver double-digit recall improvements on multi-turn agent tasks. The zero-loss consolidation across 100 audited merges suggests that judge-gated screening can prevent the silent evidence destruction that undermines existing agent memory systems. For teams building long-horizon agents where retrieval quality bounds task success, MEMOIR offers a deployment-ready architecture that updates memory priority based on actual use rather than static heuristics.