Changing where memory updates occurred shifted model retrieval accuracy
Source check: Source checked. Not independent scientific confirmation.
Finding: Across 24 update schedules with the same number of updates, a 760-million-parameter language model’s retrieval accuracy ranged from 19.7% to 34.2% on the same 6,000 prompts. The test varied which layer groups updated most often while holding the checkpoint and evaluated prompts fixed.
Technology: Memory-as-Context MAC-760M; scheduled learned memory updates
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- The learned model updated its persistent memory during the task. Controlled placement and frequency changes tested the effects of the same update count on retrieval.
- Test and result
- The authors enumerated assignments of four update periods to four fixed layer bands and compared accuracy on each of the same exact-retrieval items.
- Limitation / still unknown
- The response range applies to one checkpoint and one retrieval task. The best and worst schedules were identified retrospectively; the ranking need not transfer.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.36109v1 · https://arxiv.org/html/2610.36109v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: §§3–4, Figure 2 and Table 1.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.