EvoMemState: Evolving Memory State for Personal Agents

EvoMemState: Evolving Memory State for Personal Agents

Long-term memory is supposed to make a personal agent feel continuous rather than forgetful. But remembering isolated facts is not the same as understanding a user’s current situation. A preference may be outdated, a newer statement may supersede an older one, and a constraint may change how an otherwise relevant fact should be used.

In EvoMemState: Evolving Memory State for Personal Agents, Zhishang Xiang, Yunbo Tang, and their coauthors argue that personal-agent memory should be treated as an evolving state, not merely a collection of retrieved records. Their framework adds relational context around stored facts and learns how that context should influence a response to a particular request.

The problem with fact-centric memory

Many memory systems extract atomic facts, organize or summarize them, and retrieve records that appear relevant to a new query. This is useful, but the authors identify two failure modes. First, semantic retrieval can find one relevant fact while missing the updates, conflicts, or constraints needed to interpret it. Second, even when all relevant facts are present, the response model may give too much weight to an old preference or apply a fact outside its intended scope.

The paper’s preliminary analysis illustrates both issues. On a subset of MemSyco-Bench designed around linked memories, several evaluated systems frequently failed to retrieve all required information. Across four of the five systems studied, answer accuracy was 5.9 to 33.3 percentage points higher when all necessary memories were retrieved than when some were missing. Yet access alone was not enough: fixed self-reflection reduced average accuracy from 50.5% to 48.9% in the authors’ comparison.

This motivates a distinction between remembering what was said and determining how it should matter now.

How EvoMemState works

EvoMemState represents memory with Persistent Facts and two forms of state.

  • Relational States record a fact’s validity, scope, source-related information, and relationships to other facts. Relations include updates, supersession, constraints, and unresolved conflicts.
  • Reasoning State is generated for a particular request. It specifies which memories apply, what role they should play, and how they should be prioritized.

The first component, Ontology-Guided Memory State Induction, extracts time-ordered atomic facts and incrementally places them in a personal memory ontology. The ontology groups facts under concepts and connects them with typed relations. New information can create, merge, or split concepts, while a newer fact can supersede an older one without deleting the original record. Conflicts can remain explicit rather than being silently resolved.

The second component, RL-Based State Evolution, trains a separate Memory State Evolution Policy. For a request and its retrieved evidence, the policy generates candidate Reasoning States. Each candidate is supplied to a frozen personal agent, and its reward is based on whether the resulting answer improves over a baseline that uses the same evidence without the Reasoning State. The authors optimize the policy with GRPO, leaving the response model unchanged.

At inference time, semantic retrieval first selects seed facts. Relation traversal then expands those seeds to include connected updates, conflicts, and constraints. The policy generates request-specific guidance, and the personal agent answers using the facts, their Relational States, and the Reasoning State together.

What the experiments show

The authors evaluate the system on MemSyco-Bench, MemTrapBench, LoCoMo, and LoCoMo-Refined, using Qwen3-8B and GPT-4o-mini as response models. The benchmarks cover long-term recall, preference changes, memory-induced sycophancy, evidence priority, safety, and task-boundary errors.

On MemSyco-Bench, EvoMemState reaches 57.6% accuracy with Qwen3-8B and 58.9% with GPT-4o-mini. The authors report that these results exceed the strongest baselines by 12.2 and 5.9 percentage points, respectively. With Qwen3-8B, the system also records the best results for deciding when and how memory should be used, the lowest reported sycophancy rate, and the highest Correct Use score in that evaluation.

On LoCoMo-Refined, EvoMemState achieves the highest LLM-judge and token-F1 scores for both response models. Its judge score exceeds the strongest baseline by 6.1 percentage points with Qwen3-8B and 7.0 points with GPT-4o-mini. On MemTrapBench, the authors report the best Safety and Task Boundary results in their comparison.

The component studies support the proposed division of labor. With Qwen3-8B, removing Relational States reduces MemSyco-Bench accuracy by 8.9 points and the MemTrapBench score by 6.4 points. Removing the Reasoning State reduces those results by 8.0 and 15.0 points. The authors also report a 93.7% complete-retrieval rate for EvoMemState on a valid-memory-selection evaluation; among those complete-retrieval cases, adding a Reasoning State raises accuracy from 75.4% with self-reflection to 86.7%.

There is also an efficiency signal, although it should not be treated as a universal guarantee. On LoCoMo-Refined, the system retrieves an average of 565 tokens and reaches a retrieval hit rate of 0.816, compared with 816 tokens and 0.789 for MemoryBank, and 1,440 tokens and 0.789 for A-MEM. These figures describe the authors’ setup and metrics, not an established general law about memory cost.

Limitations and practical implications

The most explicit limitation is scope: the study focuses on text-based interaction histories. The authors note that real personal agents must also handle documents, images, and other multimodal artifacts. Extending the ontology and its relations across modalities remains future work.

The results are also benchmark-based, and the policy is trained using answer-quality feedback from constructed question–memory instances. Performance may therefore depend on the quality of the ontology extraction, relation labels, reward evaluator, and benchmark distribution. Relation traversal can improve completeness, but it may also expand context or propagate an incorrectly extracted relation. The reported gains demonstrate the value of the proposed design in the tested settings; they do not by themselves establish reliability in unrestricted personal use.

For practitioners, the paper suggests a useful architectural principle: separate memory storage, memory relationships, and memory application. Preserve original evidence, represent supersession and conflict explicitly, and make request-specific applicability visible to the response model. A lightweight prototype need not reproduce every training detail. It could begin with typed memory records and provenance, add relation-aware retrieval, and evaluate separately whether evidence was found and whether it was used correctly.

The broader lesson is that personal-agent memory is not only a retrieval problem. It is also a state-management and decision problem. EvoMemState’s evidence suggests that agents benefit when they learn not just which memories to fetch, but when those memories should guide, constrain, or stay out of the answer.


Sources

AI-assisted research summary. Check the original paper for complete methods, results, and limitations.

Avatar photo

Tech News

Jeff is a passionate blog writer who shares clear, practical insights on technology, digital trends and AI industries. With a focus on simplicity and real-world experience, his writing helps readers understand complex topics in an accessible way. Through his blog, Jeff aims to inform, educate, and inspire curiosity, always valuing clarity, reliability, and continuous learning.

↑ ↓ navigate Enter open Esc close