Janus Controller Improves LLM Memory by Selective Updates.

Researchers have introduced Janus, a novel plug-in memory controller designed to enhance the performance of Large Language Models (LLMs) with sequentially evolving memory. Existing LLM memory systems often accept every locally generated update without verifying its long-term benefit, which can lead to the loss of valuable knowledge, the introduction of overly specific rules, or a bias towards recent examples. Janus addresses this by acting as a gatekeeper, deciding whether to accept a new memory update or retain the previous state. To make these decisions efficiently, Janus employs a Memory Momentum Trigger to detect suspicious deviations in the memory-update trajectory. It then compares the old and new memories using a compact hybrid evaluation set that assesses coverage, boundary conditions, and fresh tasks, rather than replaying the entire history. This method-agnostic controller can wrap around existing memory updaters without altering their core rules. Across six datasets, two backbone LLMs, and two memory updaters, Janus consistently improved average accuracy by 2.7 to 4.6 points over the base updaters, demonstrating its effectiveness in maintaining high-quality, stable LLM memory.