Told Once: One Memory Entry per Name in a Frozen Language Model

A fact told to a language model after its training can be kept in the prompt, written into its trained weights, or added to a memory beside them. We add it to a lookup memory the model already has: DeepSeek-V4.1-Flash carries one indexed by hashed n-grams, whose retrieved vector is added to the residual stream through a learned gate. Each name gets one entry in that memory, keyed by the name's exact tokens; its value is found from a fitted start by a few damped minimum-norm Gauss–Newton steps on conditions built from the model's own replies about people it already knows. Linear reads at the end of the question admit the entry only for a kind of question the name was told. No trained weight changes: an entry adds 40,960 new weights to the model's memory (a key and a value of 20,480 numbers each; only the value is solved for), computed in about 50 seconds and kept in a file the model loads at start. We told 120 made-up people a birthplace, a university and an occupation each, and asked every question with no told fact in the prompt and the model's reasoning switched off. The model answers the told facts for 360 of 360 direct questions, which are conditions of the solve, and for 587 of 600 in five wordings that are in no condition. It answers one-step consequences that are written nowhere for 481 of 600 (the untouched model: 102) with one value for all of a name's facts, and for 582 of 600 when the entry holds a value for each kind of fact and the read at the end of the question chooses it. In both cases the 2,520 replies about things the names were not told, and the 500 replies about real people, are identical to the untouched model's. We report the faults found on the way and the limits: three kinds of fact, a small set of values, short answers, question forms seen during development (so this is not a blind test), one step of consequence, and a first run on one of the field's benchmarks, where the two-hop question is answered for 57 of 151 edited entities (7 untouched) and the edited entity's other answers are disturbed. Told the way the made-up people are, with the model's own examples of the new value and a read made on the spot for the edited relation, fitted on the benchmark's own question wordings asked of those examples, the same cases come out 116 of 151, and the entity's other replies are unchanged for 440 of 453 questions that are in no read. On a second benchmark, run with the first form of the telling only, rewordings of a taught sentence are answered for 136 of 1,000.

Authors

Publication Details

Published
2026-11-14
DOI
https://doi.org/10.67347/pith.preprint.2026.000072
Primary Topic
Topic Modeling
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Told Once: One Memory Entry per Name in a Frozen Language Model

Jonathan Washburn
Topic Modeling
preprint

Told Once: One Memory Entry per Name in a Frozen Language Model

Jonathan Washburn
preprint en

Abstract

A fact told to a language model after its training can be kept in the prompt, written into its trained weights, or added to a memory beside them. We add it to a lookup memory the model already has: DeepSeek-V4.1-Flash carries one indexed by hashed n-grams, whose retrieved vector is added to the residual stream through a learned gate. Each name gets one entry in that memory, keyed by the name's exact tokens; its value is found from a fitted start by a few damped minimum-norm Gauss–Newton steps on conditions built from the model's own replies about people it already knows. Linear reads at the end of the question admit the entry only for a kind of question the name was told. No trained weight changes: an entry adds 40,960 new weights to the model's memory (a key and a value of 20,480 numbers each; only the value is solved for), computed in about 50 seconds and kept in a file the model loads at start. We told 120 made-up people a birthplace, a university and an occupation each, and asked every question with no told fact in the prompt and the model's reasoning switched off. The model answers the told facts for 360 of 360 direct questions, which are conditions of the solve, and for 587 of 600 in five wordings that are in no condition. It answers one-step consequences that are written nowhere for 481 of 600 (the untouched model: 102) with one value for all of a name's facts, and for 582 of 600 when the entry holds a value for each kind of fact and the read at the end of the question chooses it. In both cases the 2,520 replies about things the names were not told, and the 500 replies about real people, are identical to the untouched model's. We report the faults found on the way and the limits: three kinds of fact, a small set of values, short answers, question forms seen during development (so this is not a blind test), one step of consequence, and a first run on one of the field's benchmarks, where the two-hop question is answered for 57 of 151 edited entities (7 untouched) and the edited entity's other answers are disturbed. Told the way the made-up people are, with the model's own examples of the new value and a read made on the spot for the edited relation, fitted on the benchmark's own question wordings asked of those examples, the same cases come out 116 of 151, and the entity's other replies are unchanged for 440 of 453 questions that are in no read. On a second benchmark, run with the first form of the telling only, rewordings of a taught sentence are answered for 136 of 1,000.

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