The Past Is Prologue: A Plug-in Controller for Selective Updates in Sequentially Evolving LLM Memory
06:00 · July 1, 2026 · arXiv cs.AI RSS
Sequentially evolving LLM memory enables agents to reuse past experience, but existing systems usually deploy each locally generated memory update without checking whether it improves future behavior. As a result, updates that help the current task may overwrite useful knowledge, introduce over-specific rules, or bias the final memory toward recent examples. We propose Janus, a plug-in memory controller that decides whether to accept a candidate memory update or retain the previous memory. To make this decision efficient, Janus uses a Memory Momentum Trigger to identify suspicious deviations in the memory-update trajectory, and compares old and new memories on a compact hybrid evaluation set of coverage, boundary, and fresh tasks instead of replaying the full history. Janus is method-agnostic and wraps existing updaters without changing their update rules. Across six datasets, two backbone LLMs, and two memory updaters, Janus improves average accuracy by +2.7 to +4.6 points over the corresponding base updaters.
Summary
Sequentially evolving LLM memory lets agents accumulate experience across tasks by updating an external store after each interaction, yet most existing systems apply every locally generated revision without verifying its effect on future performance. This produces final memories that can overwrite earlier knowledge, embed overly narrow rules, or tilt toward the most recent examples. Janus addresses the issue by acting as a lightweight, method-agnostic controller that wraps any base updater and decides whether to install a candidate memory or keep the prior version.
The controller relies on two mechanisms to keep validation tractable. A Memory Momentum Trigger monitors the trajectory of successive updates and initiates explicit comparison only when a new candidate deviates markedly from recent change directions. When comparison occurs, Janus evaluates the old and new memories on a compact hybrid batch that mixes coverage tasks drawn from the observed distribution, boundary tasks where prior memory choices altered correctness, and a small slice of recent tasks. This avoids replaying the entire history while still guarding against overfitting to stale examples.
Empirical results across six datasets, two backbone LLMs, and two different memory updaters show that the selective deployment policy raises average accuracy by 2.7 to 4.6 points relative to the unmodified base updaters. The gains stem from fewer harmful overwrites rather than any change to the underlying update rules themselves.
Why it matters
This research is highly relevant for AI researchers and developers in the Netherlands working on autonomous agents and LLM memory systems. It provides a novel, method-agnostic approach to improving the reliability and efficiency of continuous learning in LLMs, aligning with the Dutch focus on robust and high-performing AI development.
