Useful Memories Become Faulty When Continuously Updated by LLMs

TL;DR

This paper investigates how continuous memory consolidation by LLMs can degrade memory accuracy, highlighting the importance of raw episodic data and controlled consolidation for reliable agent memory.

Contribution

It reveals that current LLM-based memory consolidation often introduces faults, and proposes that raw episodes should be preserved and consolidation should be explicitly managed.

Findings

Memory utility peaks then degrades with consolidation.

GPT-5.4 fails on 54% of previously solved problems after consolidation.

Preserving raw episodes and controlling consolidation improves accuracy.

Abstract

Learning from past experience benefits from two complementary forms of memory: episodic traces -- raw trajectories of what happened -- and consolidated abstractions distilled across many episodes into reusable, schema-like lessons. Recent agentic-memory systems pursue the consolidated form: an LLM rewrites past trajectories into a textual memory bank that it continuously updates with new interactions, promising self-improving agents without parameter updates. Yet we find that such consolidated memories produced by today's LLMs are often faulty even when derived from useful experiences. As consolidation proceeds, memory utility first rises, then degrades, and can fall below the no-memory baseline. More surprisingly, even when consolidating from ground-truth solutions, GPT-5.4 fails on 54% of a set of ARC-AGI problems it had previously solved without memory. We trace the regression to the consolidation step rather than the underlying experience: the same trajectories yield qualitatively different memories under different update schedules, and an episodic-only control that simply retains those trajectories remains competitive with the consolidators we test. In a controlled ARC-AGI Stream environment that exposes Retain, Delete, and Consolidate actions, agents preserve raw episodes by default and double the accuracy of their forced-consolidation counterparts; disabling consolidation entirely (episodic management only) matches this auto regime. Practically, robust agent memory should treat raw episodes as first-class evidence and gate consolidation explicitly rather than firing it after every interaction. Looking forward, reliable agentic memory will require LLMs that can consolidate without overwriting the evidence they depend on.