Existing memory systems for large language model (LLM) agents primarily update information upon arrival, treating retrieval as a static endpoint. This approach rarely utilizes retrieval feedback to actively reorganize memory structures for future access.
The new framework, named REALM, is inspired by memory reconsolidation in cognitive neuroscience. It models long-term memory as a continual lifecycle that autonomously organizes data into a heterogeneous cognitive graph.
REALM retrieves evidence through adaptively composed graph-search atoms and reconsolidates memories based on retrieval feedback. This mechanism allows the agent to progressively reorganize related memory units into coherent local structures.
The system demonstrated an average accuracy of 75.97% on the LoCoMo benchmark and 65.11% on LongMemEval. These results represent improvements of 7.17 and 1.31 points respectively over the strongest baselines tested.
Ablation studies confirm that memory reconsolidation consistently boosts performance. The findings suggest retrieval-driven reconsolidation is an effective mechanism for evolving long-term memory in LLM agents operating over extended interactions.
Source: https://arxiv.org/abs/2609.16053
