Personal agents learn how to use evolving memories
EvoMemState: Evolving Memory State for Personal Agents · arXiv:2611.03743
In plain words
A personal agent can remember the right facts and still answer poorly if it cannot tell which facts are current, how they fit together, or which matter now. EvoMemState is a memory framework that keeps facts with links showing updates and conflicts, then adds guidance for each new request.
On LoCoMo-Refined, a test of recalling dialogue details, it scored above the strongest comparison system with Qwen3-8B. The evaluator score reflects whether answers cover reference details, preserve timing, avoid contradictions, and avoid unsupported additions. This result applies to this test and answer model.
The research
Remembered details can change or conflict, and a personal agent may need to decide which detail fits a particular request.
This work makes that decision part of the memory system: it tracks how facts relate and tests whether learned, request-specific guidance helps the agent use them more appropriately.
EvoMemState links personal facts to their changing relationships and learns to generate request-specific guidance about how an agent should use them.
EvoMemState was compared with nine memory systems on four tests of memory use and dialogue recall, using Qwen3-8B and GPT-4o-mini.
On LoCoMo-Refined, an evaluator checks coverage of reference details, timing, contradictions, and unsupported additions; with Qwen3-8B, EvoMemState scored higher than the strongest baseline, SimpleMem.
That result supports an advantage on this refined recall test. A separate MemTrapBench comparison also found a higher aggregate judge score for the full system than for a version whose guidance policy had not been trained with reinforcement learning.
The comparisons cover four text-based memory tests and two answer models. The reported advantage on LoCoMo-Refined is for Qwen3-8B, and the policy-training comparison is an aggregate on MemTrapBench with that model; the study does not cover memories drawn from images or documents.
What the evidence shows
- Facts need context
EvoMemState is a long-term memory framework for personal agents. It addresses a gap in fact-centered systems: individually stored facts can become outdated, conflict, or apply only in certain circumstances.
Even when the relevant facts are retrieved, the agent still has to decide how they should shape the current answer.
- Relations preserve change
The system keeps each original fact and adds a record of its current validity, scope, and relations to other facts. New information can update those relations without erasing earlier records. This makes updates and unresolved conflicts available as context when the agent uses memory.
- More evidence was not enough
Complete retrieval alone did not guarantee accurate use in a preliminary test. On a subset where all relevant memories were available, average answer accuracy across evaluated systems was 50.5% with direct reasoning and 48.9% with a fixed prompt asking the agent to reflect.
The comparison is limited to that subset; it points to the need for guidance fitted to each request.
- Guidance improves recall score
For each request, EvoMemState generates guidance about which memories apply and how they should shape the answer, while keeping the underlying facts available to the response model.
On LoCoMo-Refined, a dialogue-recall test whose evaluator checks reference coverage, timing, contradictions, and unsupported details, it scored above nine memory-system baselines with Qwen3-8B, including the strongest baseline, SimpleMem.
- Training matters in one test
A separate MemTrapBench comparison helps assess the learned guidance mechanism. With Qwen3-8B, the overall judge score was 79.1 for the full system and 62.7 for a version that generated guidance with the policy before reinforcement-learning training.
This supports the role of policy training in that test setup, while the score is an aggregate across its four scenario types.
Deep
- Original paper authors
- Zhishang Xiang, Yunbo Tang, Zhongqiao Liu, Junkai Chen, Zerui Chen
- Submitted to arXiv
- License
- CC BY 4.0