Long‑term interaction of LLM agents requires a reliable memory. Existing systems only update memory when new information arrives and treat retrieval as the final step, missing the opportunity to use retrieval feedback for continuous memory reorganization. Most approaches rely on predefined memory structures and fixed retrieval pipelines, limiting an agent’s ability to autonomously structure its memory. Inspired by memory reconsolidation in cognitive neuroscience, we propose REALM (reconsolidation‑evolution agentic long‑term memory), which models memory as an ongoing lifecycle. The framework automatically arranges memories into a heterogeneous cognitive graph, retrieves evidence through adaptively composed graph‑search atoms, and continuously reconsolidates memories based on retrieval feedback. Experiments achieve 75.97% average accuracy on LoCoMo and 65.11% on LongMemEval, surpassing the strongest baselines by 7.17 and 1.31 points respectively. Ablation studies confirm that reconsolidation consistently improves performance, and further analysis shows it progressively reshapes related memory units into tighter local structures for collective evidence recall during reasoning. These findings indicate that retrieval‑driven memory reconsolidation offers an effective mechanism for the continual evolution of long‑term memory in LLM agents.
Review