STARK paper proposes dual-graph memory for personalized AI agents
A new arXiv paper introduces STARK, a memory architecture that links user events to behavioral insights, aiming to help agents learn about individuals over time.
In plain words
AI agents that work with people over long periods need memory, but most current designs only record what happened and leave the interpretation to the language model at query time. STARK is a proposed research architecture that stores events in one graph and the behavioral lessons drawn from them in another, connecting the two. The authors report score improvements on memory benchmarks, including for new users with little history.
What was announced
The paper, published on arXiv, argues that current agent architectures capture what happened in a user's interactions while the reasoning about why it happened is left to the underlying large language models to reconstruct later. STARK instead uses two connected graphs. An Episodic Graph records interactions as episodes linked across time, cause, and meaning, while an Insight Graph holds behavioral signals extracted by an LLM, scored by how many events support them. Events themselves serve as edges connecting to the insights they produced.
At inference time, the system retrieves behavioral knowledge directly from insight nodes, which the authors say avoids repeatedly walking through episode history. Because events act as edges, the graph can be projected into a relational embedding space, allowing knowledge transfer between users who share similar contexts and helping infer missing insights for new users with no history. The reported results include a 1.48 judge-score gain on LongMemEval information extraction and 0.93 on LoCoMo open-domain questions over the stated state of the art, plus a 0.7-point judge-score and 3-point ROUGE-L improvement on LongLaMP with cross-user transfer at 5% graph sparsity. These figures come from the authors' own evaluations and have not been independently verified.
What it does not show
- The source is the paper's abstract only; full method details, experimental setup, and baselines are not covered in the supplied excerpt.
- Reported benchmark gains are the authors' own claims and are not independently confirmed.
- The abstract does not describe a deployed system, implementation availability, or real-world usage results.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.