Dettmers releases CliffCompaction: 50% token savings for million-token coding agent sessions
Tim_Dettmers · x · 2026-09-23
Tim Dettmers' team open-sourced CliffCompaction, an autocompaction tool for coding agents that supports million-token sessions and cuts token costs by up to 50% while matching or beating baselines on Terminal-Bench and KernelBench.
- Key principle: only truncate or drop content, never rephrase; never compact a compaction—each pass operates on original content only, preventing context drift.
- Results: under parallel test-time scaling, Kimi K2.6 + CliffCompaction matches Opus 4.7 and beats Opus 4.6 and GPT-5.3 Codex at lower cost; +10pp on Terminal-Bench for less than two full-context runs.
- Continual learning: reaches 2.23x CUDA kernel speedup at 200 steps and 3.58x at 400 steps on KernelBench, surpassing specialized search algorithms and trained agents.
- Integration: works with Claude Code and Codex; the team says it has better vibes than their built-in compaction after months of use. Paper, code, PyPI and blog are all available.
More from coding & agent
- DeepLearning.AI launches free course on building on-device memory AI assistants with Qdrant — DeepLearningAI · 2026-09-23
- Hugging Face ships lerobot JS package to read robot datasets in the browser — mishig25 · 2026-09-23
- DavidKPiano: a coding agent's "2-3 days" estimate means days fixing its missed edge cases — DavidKPiano · 2026-09-23
- Sol 6 agent still autonomously running on a single goal after 21 hours — gregmushen · 2026-09-23
- Harvard, Andrew Ng and Karpathy Teach AI Engineering Free: A 9-Step Learning Path — mdancho84 · 2026-09-23
- Cookie banners vs agents: X user shows Europe's web friction makes AI agents look superhuman — jeff_weinstein · 2026-09-23