Smarter Compact open-sourced: 72% cheaper agent sessions, nothing lost, all searchable
goost95 · reddit · 2026-11-06
The same author's introduction post for Smarter Compact: local SQLite stores the full conversation, old turns are trimmed at the token limit, and MCP tools let the agent search, run SQL, or page back trimmed content — zero LLM summarization calls. ContextBench with Haiku 5.5: 0.35→1.00 score on the longest task, peak prompt 762k→95k, $6.07→$0.38 per trial, 72% average savings. Lessons: trimming to half the limit beats cutting to 30%; trimmed user turns may contain corrections the note should flag.
Related event: Smarter Compact Cuts Long Agent Session Costs 72%(2 posts)→
More from coding & agent
- Scyvera open-sources agent governance: scan repos, score blast radius, auto-draft contracts — Trout_dev · 2026-11-06
- How do you catch an agent action that succeeded but was wrong? — krunal_builds · 2026-11-06
- RAME Code: open-source VS Code coding agent runs local Muse Glimmer 30B with permission-controlled tools — acaimari · 2026-11-06
- Open-source tjcoder/cli: terminal agent platform where the model directs its own UI — tjcoder-labs · 2026-11-06
- GitHub rebuilds Git storage for thousands of concurrent AI agents — arpit_bhayani · 2026-11-06
- OpenAI built Pando for Codex: 5x lower Git push latency, ~1PB transferred daily — sven_ai · 2026-11-06