Researchers from HKUST (Guangzhou), Jilin University and the Chinese University of Hong Kong have proposed HeLa-Mem (Hebbian Learning Associative Memory), which targets the gap where an AI agent's relevant memories share little wording, and the paper was accepted to ACL 2026 Main with code published on GitHub. Each dialogue turn becomes an episodic memory node; links between nodes strengthen when they are retrieved together and decay when unused, while a reflection agent distils reusable semantic memory from densely connected nodes and flags isolated, stale nodes as forgetting candidates. Retrieval runs both a similarity/keyword/time path and an association-spreading path, merging results with semantic memory for the LLM; on LoCoMo with GPT-4o-mini, the authors report F1 of 40.14 (multi-hop), 47.29 (temporal), 29.70 (open-domain) and 51.89 (single-hop), using about 1010 retrieval-context tokens on average, versus roughly 2000 cited for MemoryOS and about 16910 for full-dialogue input. The authors note association weights must accumulate through interaction, leaving a cold-start challenge when history is thin, and that the distilled knowledge depends on the base LLM's capability.