Skip to content

[Idea] A fixed-footprint personalization substrate as a memory layer for on-device agents #516

Description

@igs0707

Hi RunAnywhere team,

Saw your YC launch post asking what sucks about current on-device AI
stacks — here's a data point, not a feature request.

I built a small (N=12) reference architecture for a memory substrate that
stays a genuinely fixed size (kilobytes, not a growing log) no matter how
long an agent runs — verified across a 256x range of lifetime length, with
recall quality for early memories barely moving (0.81 to 0.80). The same
architecture individuates per user through lived experience alone: an
agent that lived with one user measurably outperforms one trained on a
different user, and a "stranger" agent can perform worse than a blank one.

This could be a natural memory-layer complement to RunAnywhere's inference
stack — full per-user personalization with no growing storage, fitting the
same memory-pressure constraints your SDK already manages for models.

Preprint + code: https://doi.org/10.5281/zenodo.21122080

Sharing in case it's a useful reference point — not pitching anything,
just curious if this is a direction worth exploring.

— Kimiyasu Igarashi, independent researcher

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions