EvoMemState gives AI memories context—and rules for using it

Remembering facts is only the beginning

A personal AI assistant can retrieve an accurate statement and still give an inappropriate answer. An old preference may have been replaced, a remembered belief may conflict with objective evidence, or information relevant to one situation may not apply to another.

In the November 2, 2026 arXiv preprint EvoMemState: Evolving Memory State for Personal Agents, Zhishang Xiang and colleagues propose separating these problems. Their framework preserves factual records, explicitly tracks relationships between them, and learns to generate guidance about using those records for a particular request. The reported results concern benchmark performance, not demonstrated reliability in everyday deployment.

Keeping the past without letting it govern the present

EvoMemState starts by extracting atomic records from dialogue: individual propositions with source information, timestamps and qualifiers. These “Persistent Facts” retain their original content even when later conversations change their significance.

The system organizes records beneath a hierarchy of concepts and attaches “Relational States” describing their validity, scope and relationships. A newer preference can supersede an older one without deleting it. Contradictory records can remain explicitly unresolved instead of being merged into an apparently consistent summary.

This distinction preserves historical information. A former preference might be unsuitable for a current recommendation but essential when answering a question about the user's past interests.

For retrieval, the implementation first selects ten records by embedding similarity. It then follows stored memory relations in both directions to collect connected records, retaining their relationship types and removing duplicates. Consequently, the final evidence set can extend beyond the initial ten matches.

Learning how memories should influence an answer

Relationships alone do not determine what to do with evidence. EvoMemState therefore adds a request-specific “Reasoning State”: compact natural-language guidance identifying which memories apply, their roles and their relative priority.

The paper illustrates this with a vacation scenario involving an earlier hiking preference, a newer beach preference and a knee injury. For a current recommendation, the old preference supplies historical context, the newer one guides personalization, and the injury constrains suitable activities. A historical question would require different treatment of the same records.

A separate policy model learns to produce this guidance through reinforcement learning. During training, a frozen response model answers using the same evidence with and without candidate guidance. The answer-score difference supplies a reward, with additional checks for formatting and reference-answer disclosure. Only the guidance policy is updated.

The authors construct approximately 60,000 training examples from FEVER, SciFact and PersonaMem, including conflicts between personal beliefs and objective evidence, preference changes and distracting memories.

What the experiments report

The evaluation compares EvoMemState with nine baselines across four benchmarks using Qwen3-8B and GPT-4o-mini as response models.

On MemSyco-Bench, which examines appropriate memory use, EvoMemState reports average accuracies of 57.6% and 58.9%, respectively. These exceed the strongest baseline averages by 12.2 and 5.9 percentage points. On LoCoMo-Refined, its answer-judge scores reach 57.4 and 60.8, compared with the strongest baseline scores of 51.3 and 53.8.

The advantages are not universal. With Qwen3-8B, EvoMemState's MemTrapBench average is 79.1, below SimpleMem's 79.9. It also does not lead every metric on the original LoCoMo benchmark.

Ablations support contributions from both kinds of state within the tested setup. Removing relational state lowers Qwen3-8B MemSyco-Bench accuracy from 57.6% to 48.7%; removing reasoning state lowers it to 49.6%. Keeping guidance generation but removing reinforcement learning produces 49.7%.

What remains uncertain

The study explicitly focuses on text-based interaction histories. Extending its approach to images, documents and other modalities remains future work.

Retrieval is also incomplete: on LoCoMo-Refined, EvoMemState recovers every required evidence item for 68.0% of questions, versus 66.9% for MemoryBank. The authors caution that accuracy measured only on complete-retrieval cases involves different question subsets across systems.

The central contribution is therefore a tested architectural distinction, not a guarantee of dependable personal assistance: preserving information, maintaining its changing context and deciding how to apply it are separate tasks.