EvoMemState: Self-Evolving Memory Paradigm for Personal Agents

Introduction

Each intelligence paradigm shift redefines the knowledge infrastructure systems require. Personal Agents need a memory with evolvable state: facts must be traceable, judgments verifiable, user state continuously updatable, and memory's role dynamically adjustable per task. Based on this insight, Xiamen University and Jilin University propose EvoMemState: Evolving Memory State for Personal Agents , moving long-term memory from Fact-Centric Storage to State-Centric Evolution .

01 Why Remembering Facts Still Leads to Errors

Existing long-term memory systems largely follow a Fact-Centric paradigm: extract user facts from dialogue, store, compress, or update them, then retrieve relevant records for the current question and feed them to the LLM. This paradigm solves the context-window limitation but rests on two assumptions that do not always hold: (1) retrieved facts suffice to reconstruct the user's current state; (2) given correct facts, the model will automatically use them correctly.

Fact relevance does not equal evidence completeness. On the MemSyco-Bench Valid Memory Selection subset, MemoryBank, LightMem, and Supermemory failed to retrieve all necessary information in 58.9%, 61.8%, and 58.0% of samples respectively. For four of the five tested systems, full evidence retrieval improved answer accuracy by 5.9–33.3 percentage points over incomplete evidence.

Evidence completeness does not equal correct usage.

02 From Fact-Centric to State-Centric: Memory = Facts + States

The authors propose a State-Centric personal memory paradigm, redefining the long-term memory infrastructure for Personal Agents: personal memory is a system co-constituted by facts and states, continuously evolving through interactions and serving current decisions. EvoMemState implements this paradigm via three synergistic layers: Persistent Facts, Relational States, and Reasoning State.

03 EvoMemState: Building, Learning, and Using State

EvoMemState consists of three sequential stages: Ontology-Guided Memory State Induction, RL-Based State Evolution, and State-Aware Reasoning.

Stage 1: Ontology-Guided Induction

Method: Ontology-Guided Induction

Role: Extract atomic facts from dialogue, build personal memory ontology, maintain validity, scope, and relations.

Stage 2: RL-Based State Evolution

Method: RL-Based State Evolution

Role: Train independent policies based on downstream answer quality, learn reusable memory usage experience.

Stage 3: State-Aware Reasoning

Method: State-Aware Reasoning

Role: Complete associated evidence, generate request-level reasoning state, then guide frozen Personal Agent to answer.

04 Why Both States Are Indispensable

Assume long-term memory holds three records: M1 "past liked hiking", M2 "now prefers beach travel", M3 "recovering from knee injury".

M1 — System stores: past liked hiking · historical. Current vacation request usage rule: M1 can answer historical interest but does not dominate current recommendation.

M2 — System stores: now prefers beach · current. Current vacation request usage rule: M2 represents current preference, should have higher priority.

M3 — System stores: knee injury recovering · constraint. Current vacation request usage rule: M3 is a reality constraint, should exclude high-intensity activities.

05 Experimental Results: Not Only More Complete Retrieval, But Also More Correct Usage

The paper evaluates on MemSyco-Bench, MemTrapBench, LoCoMo, and LoCoMo-Refined, comparing nine baselines: NaiveRAG, MemGPT, MemoryBank, Mem0, A-MEM, LightMem, Supermemory, SimpleMem, PGMem. Response models: Qwen3-8B and GPT-4o-mini.

Key Metrics

MemSyco Qwen3-8B: 57.6%

MemSyco GPT-4o-mini: 58.9%

Average Complete Retrieval Rate: 85.6%

LoCoMo-Refined Retrieval Hit Rate: 0.816

More Reliable Judgment of When and How to Use Memory

EvoMemState achieves 57.6% and 58.9% average accuracy on MemSyco-Bench, surpassing the strongest baseline by 12.2 and 5.9 percentage points. Under Qwen3-8B it simultaneously attains the lowest memory sycophancy rate and highest correct memory usage rate.

Complete Retrieval Is Just the Start, Correct Usage Is Key

EvoMemState's average Complete Retrieval Rate reaches 85.6%. On evidence-complete samples, LoCoMo-Refined Acc@complete reaches 70.8% (+11.5 pp over strongest baseline); MemSyco-Bench Acc@complete reaches 68.5% (+31.2 pp).

State Modeling Does Not Sacrifice Long-Term Fact Recall

Qwen3-8B — MemSyco: 57.6, MemTrap: 79.1, LoCoMo-R: 57.4, Overall Avg: 59.9

GPT-4o-mini — MemSyco: 58.9, MemTrap: 78.7, LoCoMo-R: 60.8, Overall Avg: 61.2

On LoCoMo-Refined, EvoMemState achieves the highest LLM-Judge score and Token F1 with both response models, Judge Score exceeding the strongest baseline by 6.1 and 7.0 percentage points. Preserving raw facts and explicitly maintaining relations means state reasoning does not trade off long-term recall ability.

https://arxiv.org/pdf/2611.03743
https://github.com/XMUDeepLIT/EvoMemState