Observed arrival · 2026-11-18
CompactBench puts AI memory loss through repeat compaction
An open benchmark tests how much information AI context-compaction methods retain after five rewrites under a fixed 1,024-token memory budget.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
- For
- AI agent and memory-system developers
- Worth noticing
- Its example catches a method retaining Marlin’s old go-live date after compaction four, despite a later correction.
Field notes
The sample places a one-time date correction among code, logs, and figures from other projects, then shows a sectioned-summary method keeping the superseded date after compaction four. The listed failure families also test scoped rules, reference resolution, and whether work marked done is actually only in review. The page says leaderboard scores are family-balanced and include a 95% interval.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue