Agent Zero Memory: Provenance-Aware Long-Term Memory for LLM Agents
Agent Zero Memory splits agent memory into three cited stores to reduce blind spots and force abstention over guessing.
The paper describes a system that turns conversations, files, and connected sources into an event timeline, an entity-event knowledge graph, and a curated documentary memory. Retrieval uses an intent gate, routes to the right sources, searches the three systems in parallel, then merges grounded answers with one confidence score. Its central constraint is provenance: learned items carry origin, timestamp, and evidence, and answers can cite only material the reader actually opened. The authors report 95.60% on LongMemEval and 93.60% on LoCoMo, plus a study suggesting memory design can preserve quality while cutting query cost across different backbone models. ArXiv · AI/CL/LG's note
The paper describes a system that turns conversations, files, and connected sources into an event timeline, an entity-event knowledge graph, and a curated documentary memory. Retrieval uses an intent gate, routes to the right sources, searches the three systems in parallel, then merges grounded answers with one confidence score. Its central constraint is provenance: learned items carry origin, timestamp, and evidence, and answers can cite only material the reader actually opened. The authors report 95.60% on LongMemEval and 93.60% on LoCoMo, plus a study suggesting memory design can preserve quality while cutting query cost across different backbone models. ArXiv · AI/CL/LG's note
score 5