AuthorsKaichao Liang, Yuqi Cui, Hao Kong et al.
TL;DR
MindMemOS uses an entity–property–time memory OS with MindMemEvolve and dreaming to reach 94.03% LOCOMO accuracy and 70.63% PersonaMem accuracy.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing memory systems remain largely fixed after deployment, limiting adaptation of memory models, organization strategies, and procedural knowledge over time.
This rigidity leaves systems like EverOS at 93.05% LOCOMO accuracy and 67.57% PersonaMem accuracy, with redundant, outdated, and conflicting memories degrading long-term personalization, reasoning, and skill reuse.
HOW IT WORKS
MindMemOS combines a 3D entity–property–time memory structure with MindVanilla, MindSchema, MindMemEvolve, dreaming, feedback, and MindSkillEvolve to manage the full memory lifecycle.
You can think of MindMemOS as an OS where the entity–property–time graph is disk, compact search is the index, and MindMemEvolve plus dreaming are background defrag and schema-tuning daemons.
This design lets MindMemOS adapt schemas, consolidate conflicts, and turn trajectories into skills, going far beyond what a static context window or fixed vector store can express.
DIAGRAM
This diagram shows how MindMemOS performs hybrid bidirectional compact search over the entity–property–time graph for a query.
DIAGRAM
This diagram shows how MindMemOS is evaluated on dialogue memory, personalization, consolidation, and skill evolution benchmarks.
PROCESS
Memory-Add MindVanilla
MindMemOS uses MindVanilla to segment turns, retrieve related memories, and extract flat records with coarse types before writing into the entity–property–time graph.
Memory-Add MindSchema
MindMemOS applies MindSchema to map dialogue to scenario-adaptive entities and properties, build an episodic memory subgraph, and perform entity fusion and graph merge.
Compact Search
MindMemOS runs compact search with hybrid BM25 and embedding retrieval plus forward and reverse traversal over the 3D memory structure to gather relevant evidence.
Dreaming and Feedback
MindMemOS periodically runs dreaming to consolidate conflicts and redundancy, and uses explicit and implicit feedback to add, update, or delete memories and drive MindMemEvolve and MindSkillEvolve.
KEY CONTRIBUTIONS
Scenario-Adaptive Memory Modeling
MindMemOS introduces a unified entity–property–time structure with MindVanilla and MindSchema, achieving 94.03% overall accuracy on LOCOMO and 70.63% on PersonaMem.
Active Memory Pattern Discovery
MindMemOS uses MindMemEvolve, an LLM-guided evolutionary search over schemas, to add first-order and higher-order properties that raise PersonaMem test accuracy from 61.07% to 64.63%.
Experience-Driven Skill Evolution
MindMemOS introduces MindSkillEvolve to turn execution trajectories into versioned skills, improving SpreadsheetBench success from 48.0% with Init-skill to 57.2% with MindSkillEvolve-Sup.
RESULTS
Overall accuracy
94.03%
+0.98 over EverOS on LOCOMO
PersonaMem overall
70.63%
+3.06 over EverOS
FactConsolidation overall
0.585
+0.075 over Infini Memory with gpt-5-mini
SpreadsheetBench success
57.2%
+9.2 over Init-skill baseline
MindMemOS is evaluated on LOCOMO, PersonaMem, MemoryAgentBench FactConsolidation, and SpreadsheetBench, showing consistent gains in recall, personalization, consolidation, and skill execution. The 94.03% LOCOMO and 70.63% PersonaMem scores demonstrate that MindMemOS’s evolving memory OS can match or exceed specialized baselines across diverse long-term tasks.
BENCHMARK
MindMemOS is evaluated on LOCOMO, PersonaMem, MemoryAgentBench FactConsolidation, and SpreadsheetBench, showing consistent gains in recall, personalization, consolidation, and skill execution. The 94.03% LOCOMO and 70.63% PersonaMem scores demonstrate that MindMemOS’s evolving memory OS can match or exceed specialized baselines across diverse long-term tasks.
BENCHMARK
Overall accuracy (%) on LOCOMO across memory systems.
BENCHMARK
Overall accuracy (%) on PersonaMem across memory systems.
KEY INSIGHT
Dreaming in MindMemOS archives roughly 22.5% of active memories yet increases FactConsolidation overall accuracy from 0.545 to 0.585 with gpt-5-mini.
This is surprising because aggressively compressing memory is usually expected to hurt retrieval, but MindMemOS shows that targeted consolidation can both shrink and sharpen long-term stores.
WHY IT MATTERS
MindMemOS makes it practical to run agents with a self-evolving, schema-aware memory OS that refines both content and skills over time.
Builders can now plug MindMemOS behind diverse agents and get portable long-term memory, automatic schema tuning, and trajectory-driven skill upgrades without redesigning their runtimes.
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…