AI memory tested: GPT-4o drops from 87% to 60.6% when fed the full 115k-token chat history
gaganghotra_ · x · 2026-10-13
A beginner's guide article on how AI memory actually works includes a telling test: GPT-4o answered 87% of questions correctly when shown only the relevant slice of a chat history, but accuracy fell to 60.6% when it saw the entire 115k-token history.
The takeaway: stuffing full context into a model isn't better memory — effective memory systems require retrieval and filtering rather than brute-force context.
More from coding & agent
- Swapping only the harness moves a model 35 points: 8090 Labs tests 7 models across 5 coding agents — pstAsiatech · 2026-10-13
- Meta launches Muse Wonder frontier model family, installable with one curl command — jiayi_pirate · 2026-10-13
- Harness code becomes a research battleground: CMU rewrites it with RL, Apple says it doesn't matter — burkov · 2026-10-13
- FrontierCode 1.1 leaderboard: Claude Opus 5.5 tops at 52% as leading models fail 40%+ of tasks — ArtificialAnlys · 2026-10-13
- Anaconda adds Kilo Agent Swarms and Enkrypt AI autonomous red-team agents — anacondainc · 2026-10-13
- Open-source agent workspace Open Dots adds Meta's new Muse Wonder models — matchaman11 · 2026-10-13