AuthorsMing Wu, Pengyuan Zhu
TL;DR
Agent Zero Memory uses a three-store, provenance-aware, citation-locked memory substrate to reach 95.60% on LongMemEval and 93.60% on LoCoMo.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Many existing memory systems overwrite or delete facts, so corrections silently erase history and provenance is discarded when updates occur.
This breaks knowledge-update and temporal questions, where agents must know what changed and when, causing failures on benchmarks like LongMemEval and LoCoMo.
HOW IT WORKS
Agent Zero Memory centers on a triple store of Memory Events timeline, Ontology Graph, Hierarchical Documentary Memory, plus an Intent Gate and AgenticSearch-based Memory Injection.
You can think of Agent Zero Memory like RAM plus disk plus a card catalog: fast working memory, durable documentary memory, and an entity–event index that links everything.
This layered, citation-locked design lets Agent Zero Memory answer temporal, multi-hop, and update questions faithfully in ways a plain context window or flat vector store cannot.
DIAGRAM
This diagram shows how Agent Zero Memory runs Algorithm 1 to gate, route, search three memories, and integrate a citation-locked answer.
DIAGRAM
This diagram shows how Agent Zero Memory builds memory once, then runs controlled evaluations on LongMemEval and LoCoMo across eight backbones.
PROCESS
Preprocess and classify
Agent Zero Memory first runs Preprocess and classify to normalize chats, files, and sessions before feeding them into the Memory Events timeline and Hierarchical Documentary Memory.
Index
In the Index stage, Agent Zero Memory builds both embedding and lexical indexes so every artifact is searchable by meaning and exact terms.
Extract
During Extract, Agent Zero Memory uses LLM analysis to distil Memory Events and populate the Hierarchical Documentary Memory with curated, provenanced entries.
Connect
Finally, Connect builds the Ontology Graph linking entities to events, enabling multi-hop AgenticSearch during Memory Injection for complex queries.
KEY CONTRIBUTIONS
Three interconnected memory systems
Agent Zero Memory introduces a triple store of Memory Events timeline, Entity–Event Knowledge Graph, and Hierarchical Documentary Memory, built by a four-stage pipeline and queried by three concurrent AgenticSearch processes.
Provenanced items and citation-locked answers
Agent Zero Memory formalizes Provenanced items and Citation-locked answers, structurally excluding fabrication and forcing abstention when no opened evidence supports an answer.
State-of-the-art benchmark results
Agent Zero Memory achieves 95.60% on LongMemEval and 93.60% on LoCoMo, improving over Mastra by +0.73 and Mem0 by +1.10 points while remaining backbone-agnostic.
RESULTS
LongMemEval accuracy
95.60%
+0.73 over Mastra
LoCoMo accuracy
93.60%
+1.10 over Mem0
Backbone spread
3.4 points
accuracy range 92.20%–95.60% across eight backbones
Cost variation
≈30×
per-query cost range with near-state-of-the-art quality at up to 20× lower cost
LongMemEval and LoCoMo test long-term conversational memory, including temporal and knowledge-update questions, and Agent Zero Memory shows that a provenance-aware, layered memory substrate can dominate these tasks independent of backbone choice.
BENCHMARK
LongMemEval and LoCoMo test long-term conversational memory, including temporal and knowledge-update questions, and Agent Zero Memory shows that a provenance-aware, layered memory substrate can dominate these tasks independent of backbone choice.
BENCHMARK
Judged answer accuracy (%) on LongMemEval across Agent Zero Memory and strongest baselines.
BENCHMARK
Judged answer accuracy (%) on LoCoMo for Agent Zero Memory and key baselines.
KEY INSIGHT
Agent Zero Memory sees accuracy vary by only 3.4 points across eight backbones, while per-query cost varies by roughly 30×.
This challenges the assumption that bigger or more expensive models are necessary for top accuracy, showing memory-driven quality can rival model-driven scaling.
WHY IT MATTERS
Agent Zero Memory unlocks durable, provenance-aware organizational memory where every answer is citation-locked to opened evidence across timelines, graphs, and curated documents.
Builders can now deploy long-lived agents that handle knowledge updates, temporal reasoning, and multi-hop recall accurately, even with compact, cost-efficient backbones.
Jingbiao Mei, Jinghong Chen et al.
arXiv 2026 · 2026
ATM-Bench structurally evaluates long-term multimodal personal memory using Memory Ingestion, Retrieval, and Answer Generation with Schema-Guided Memory and Descriptive Memory variants. On ATM-Bench-Hard, Oracle with SGM reaches 47.3% QS while the best full system stays under 20% accuracy, revealing a large gap.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Answers use this explainer on Memory Papers.
Checking…