Beyond Retrieval: Engram and Glia for Two-Timescale Memory Regulation in Long-Horizon LLM Agents

Retrieval-Augmented Generation (RAG) can surface relevant text, but it does not manage what should update or remain available as an interaction evolves [7]. This matters: LongMemEval reports roughly a 30% accuracy drop for commercial assistants and long-context LLMs on sustained-interaction memory tasks [15]. Long-horizon LLM agents must therefore move beyond retrieval. We distinguish two competencies in long-horizon memory: trace-level selection and maintenance, and state-dependent regulation of memory engagement. We therefore introduce Engram and Glia: Engram selects episodic traces, while Glia regulates six memory operations without selecting trace identity [6, 2]. Across 140 held-out tasks with a frozen LLM policy, Engram improves full-episode success from 76.4% without memory to 90.7%, and Top-1 retrieval from 62.0% with semantic memory to 81.7%. Learned Glia closely reproduces its Synthetic regulatory target (r = .985 across phase-conditioned output profiles) and reduces retrieval events from 1,506 under Static Lifecycle control to 1,064 under Synthetic-Pretrained Glia. This is important as unnecessary retrieval can introduce irrelevant information and consume computation. Our results suggest that in long-horizon memory, a coordination problem between local memory traces and slower regulation of memory use is more sensible than a single retrieval problem. Link to code: https://anonymous.4open.science/r/memory-agent-18experiments-1C32/