Zero-Mem: Zero-Token Memory Operations for LLM Agents

Summary

Keeps original interaction traces as the source of record and builds two non-generative views: an entity–context graph (spaCy NER, co-occurrence + adjacency, Personalized PageRank) and a temporal hierarchy (turn/window/episode/local). Query routing fuses the views, then deterministic calibration filters and ranks evidence; only the final-QA reader invokes an LLM. On LoCoMo with GPT-4o-mini, average F1 59.15 vs GAM 53.75; HotpotQA 56K–448K also leads. Memory-operation LLM tokens are 0 and latency falls 57.6% vs LightMem. Official code promised after peer review.

Keywords

agent-memory · zero-token · locomo · hotpotqa · graph-memory · temporal-hierarchy · ppr · bm25 · bge-m3

Topics

LLM agents, memory systems, retrieval

Research notes

• Primary: arxiv abs (CC BY-NC-SA 4.0, cs.CL). Official repo https://github.com/TheMoon0815/Zero-mem is a placeholder until after peer review. Independent MIT Rust reimplementation exists at ptaranat/zeromem; not the authors' code. Discord posted PDF link.