How does a language-model agent remember you without sending data to the cloud?

A new preprint describes a memory system called HGP that runs entirely on a user device. It replaces the single-vector stores used by most agents with a hybrid graph that separates episodic, semantic, and procedural memory while a tiny classifier decides where each piece belongs.

Why the usual vector store breaks on a phone

Most agents today compress every interaction into one dense vector. That vector blurs the difference between a fact you stated once, a habit you repeat daily, and the exact sequence of steps you took yesterday. When the agent later searches for relevant context, the blur produces false matches and misses the implicit constraints that shape your current goal. The problem grows sharper on a phone or laptop because the model that could disambiguate those vectors lives in a data center, not in your pocket. Bandwidth and latency make round-trips impractical for real-time assistance. Local compute is limited. Battery matters.

Three graphs instead of one vector

HGP builds three distinct graphs. The episodic graph records concrete events with timestamps and participants. The semantic graph holds facts and preferences that persist across sessions. The procedural graph captures routines — ordered steps the user has demonstrated repeatedly. Each graph keeps its own topology so retrieval can follow edges that match the query type. A working-memory trajectory sits on top, representing the active state and the hidden constraints that guide the next decision. The paper does not say how large these graphs grow after months of use. No growth curves appear in the abstract.

A classifier that stays off the big model

Routing a new observation to the right graph used to require a call to the large language model itself. HGP replaces that call with a lightweight self-enhancement classifier trained on the user's own history. The classifier runs locally, so the raw traces never leave the device. Fewer cloud round-trips mean lower latency and no API bill. The authors report that this design enables on-device deployment but do not publish latency numbers for specific hardware. They also do not quantify the classifier's memory footprint.

Benchmark gap of nearly seven points

Experiments on two benchmarks show the effect. On PAL-Set solution selection HGP reaches an S-score of 35.58. The strongest baseline trails by nearly 7 points. The paper names no other metrics and does not describe the second benchmark in the abstract. Code and data are published at the repository linked from the preprint page.

What the preprint leaves open

The submission date is 7 Oct 2026. The PDF weighs 1,043 KB. Seven authors are listed. The abstract does not mention privacy audits, battery impact, or how the graphs merge when a user switches devices. Those questions wait for the full text and for independent replication.