Agent Zero Memory Splits Memory Into Three Parallel Systems, Hits New Highs
dair_ai · x · 2026-09-02
Agent Zero Memory separates what agent memory systems usually collapse into one: an episodic events timeline, an entity-event knowledge graph, and a curated documentary memory of durable facts, all running side by side.
- A retrieval turn passes through an intent gate, then a source router, then three concurrent agentic searches, one per system.
- Citation lock: every stored item carries origin, timestamp and evidence pointer, and a reply may cite only evidence its reader actually opened; when evidence is missing, the system abstains.
- Results: 95.60% on LongMemEval and 93.60% on LoCoMo, both new highs.
- The cost analysis across eight backbone models is the most useful takeaway for AI builders.
A worthwhile read if you build agents with long-term memory.
More from coding & agent
- Best Visual Guide on How LLMs Actually Work + Production Agent Engineering Course — Roger_M_Taylor · 2026-09-02
- Research: LLM agent traces reduced to 7-43 state maps — rohanpaul_ai · 2026-09-02
- Software factory workflow: Agents.md and 5 skills — Rasmic · 2026-09-02
- Benchling unveils AI stack for science, cutting R&D app dev time from months to minutes — nlarusstone · 2026-09-02
- Practical guide to Claude Code: Skills, Hooks, and Multi-Agent Teams — tom_doerr · 2026-09-02
- GUI-CC: Benchmarking Contextual Consistency of GUI World Models — Lin Fu · 2026-09-02