Zero-Mem removes LLM calls from agent memory entirely, cutting memory-operation time 57.6%

arXiv·medium signalXBlueskyLinkedInCopy link

Trending on Hacker News on August 5, Zero-Mem asks whether structured memory access needs generation at all: no step outside final question answering invokes an LLM or consumes LLM tokens, with encoder compute accounted separately. It preserves original interaction traces as the record and indexes them twice — an entity-context graph for cross-interaction connections and a temporal hierarchy for conversational locality and session state — weighing both views per query, with deterministic calibration discarding conflicting evidence before the single reader call. It matched baselines on long-memory and long-context QA while cutting memory-operation time cost by 57.6% versus the fastest compared baseline.