Bigger Context Windows Solve Fitting, Not Memory: Chroma's Context Rot Shows Accuracy Drops Early
Efficient_Joke3384 · reddit · 2026-09-08
Context windows and memory are two separate investments
- Windows now reach 1M tokens (Llama 4 Scout at 10M, Magic at 100M), but the assumption that memory stops mattering doesn't hold.
- A big window solves only the fitting problem; three issues remain: sessions still reset to zero, accuracy degrades as context grows — Chroma's context rot study across 18 models found performance dropping well before the limit, with even a single irrelevant distractor measurably hurting accuracy — and you pay to reprocess the same context every time.
- The author argues memory must improve independently: recency handling and confidence signals on retrieved memories, rather than waiting for windows to grow.
More from coding & agent
- POV: asking Gemini to fill a Google Sheet with data from Gmail — omooretweets · 2026-09-08
- Codex CLI tip: Ctrl-O copies the last response as Markdown, /copy grabs just the relevant part — gabrielchua · 2026-09-08
- herdr 0.9.0 ships the most-wanted feature: one client to control agents across all your machines — ycombinator · 2026-09-08
- Hands-on with Grok Bot's cloud agent team: hiring an AI staff without a server closet — RachelVT42 · 2026-09-08
- Should Agent Skills scripts be prewritten or generated on the fly? — dotey · 2026-09-08
- Embedflow: zero-downtime embedding model migration without 108-day backfills — Potential_Low_1183 · 2026-09-08