Are we solving LLM memory by remembering more when the real problem is knowing what to forget?
Prestigious-Tip6030 · reddit · 2026-10-24
A Reddit deep-dive argues the real LLM memory challenge isn't recall but revision.
- Core example: a coding agent learns getuser(userid), the API later changes to getuser(orgid, userid), and both versions coexist in memory — the agent must decide which to trust.
- Two approaches: update stored memory in place, or keep versions and resolve conflicts at read time. Architectures differ — Transformers/RAG preserve history, linear attention compresses it, DeltaNet updates via prediction error, BDH explores Hebbian plasticity.
- Key claim: better recall ≠ better revision; models can act on stale beliefs even when correct info is present.
- Open question: do memory benchmarks test updating without losing prior truths or disturbing unrelated memories? Is this a memory-update or a reasoning problem?
Related event: The hard problem of agent memory: knowing when to forget(2 posts)→
More from Research
- Stanford-led paper traces 'bigger is better' to data-induced competition for neurons — niloofar_mire · 2026-10-24
- Frequency drives conventionalization: higher-exposure communities conventionalize more words — abenitezburraco · 2026-10-24
- GeoWind2Plan: seconds-scale 3D urban wind prediction for energy-efficient drones — AnimaAnandkumar · 2026-10-24
- Flowing water and lava come to 3D Gaussian Splatting scenes — jonstephens85 · 2026-10-24
- Immorta Bio's AI-designed 9-amino-acid peptide SenoMune-0257 targets senescent cells — DavidmComfort · 2026-10-24
- France rebuilt its tax code as MLang, a formally verified open-source compiler — fkasummer · 2026-10-24