Memory index in the weights: a proposal for agent memory whose contents stay in files

An agent's long-term memory today is a set of files plus an index of those files. The index is either loaded into the context at every session, or kept in a retrieval system outside the model. This record proposes a third place for the index. The memory contents stay in files, unchanged. Only the index is trained into a small adapter (LoRA) on the model's weights: which memories exist, and in what situation each one is needed. Before answering, the model writes the identifier of the memory it needs, or "none", and the runtime loads that file. A first test supports the idea on a small scale. With 200 invented memories, a 4B model with the index trained in picked the needed memory for 74.0 % of requests that never mention the memory. The same model reading the index in its context picked it for 43.3 %. The adapter cost about 3 % in loss on unrelated text. The record contains the proposal and an evidence note (English originals with aligned Korean versions). The idea is under test; a comparison on a larger model is in progress and has no results yet. I searched for the same design in papers, products and public repositories and did not find it; that is a statement about the search, not a claim to be first.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-13
DOI
https://doi.org/10.5281/zenodo.24184983
Primary Topic
Topic Modeling
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Memory index in the weights: a proposal for agent memory whose contents stay in files

Sunghyeun Woo
3 links
Zenodo (CERN European Organization for Nuclear Research)
Topic Modeling
preprint

Memory index in the weights: a proposal for agent memory whose contents stay in files

Sunghyeun Woo
preprint en

Abstract

An agent's long-term memory today is a set of files plus an index of those files. The index is either loaded into the context at every session, or kept in a retrieval system outside the model. This record proposes a third place for the index. The memory contents stay in files, unchanged. Only the index is trained into a small adapter (LoRA) on the model's weights: which memories exist, and in what situation each one is needed. Before answering, the model writes the identifier of the memory it needs, or "none", and the runtime loads that file. A first test supports the idea on a small scale. With 200 invented memories, a 4B model with the index trained in picked the needed memory for 74.0 % of requests that never mention the memory. The same model reading the index in its context picked it for 43.3 %. The adapter cost about 3 % in loss on unrelated text. The record contains the proposal and an evidence note (English originals with aligned Korean versions). The idea is under test; a comparison on a larger model is in progress and has no results yet. I searched for the same design in papers, products and public repositories and did not find it; that is a statement about the search, not a claim to be first.

Zenodo (CERN European Organization for Nuclear Research)
Topic Modeling
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon