DimMem: Dimensional Structuring for Efficient Long-Term Agent Memory

DimMem introduces a lightweight long-term memory framework for LLM agents using typed memory units with explicit fields such as time, location, reason, purpose, and keywords.

The structured design improves retrieval accuracy on LoCoMo-10 and LongMemEval-S while reducing per-query token usage.

A smaller fine-tuned extractor can match or outperform larger systems like LightMem, showing strong efficiency gains.

The work targets the key trade-off in agent memory: better recall with less context cost.