PRO-LONG keeps full action logs and lifts long-horizon agents by 18 points on ARC-AGI-3
rohanpaul_ai · x · 2026-07-24
The paper argues that long-horizon agents reason better when nothing is thrown away from memory.
- PRO-LONG stores every observation, action, result, and short plan in a structured log instead of repeatedly summarizing and discarding details.
- The agent later uses ordinary tools like search and Python to recover relevant evidence from that log.
- On 25 hidden-rule games in ARC-AGI-3, the method improved the same base agents by 18 percentage points on average.
- It also matched or nearly matched stronger specialized systems while using 4.2–5.8× fewer billed tokens.
- The takeaway: long-term agent memory may work best when storage stays lossless and retrieval is programmable.
Related event: PRO-LONG Boosts Long-Horizon Agents with Programmable Memory(2 posts)→
More from coding & agent
- HeyGen adds HyperFrames to its video agent for prompt-built motion graphics — toolstelegraph · 2026-07-24
- Microsoft’s Copilot stack is finally becoming useful for workplace AI — koltregaskes · 2026-07-24
- Codex and Claude Code can already run a parent-child agent workflow end to end — huangyun_122 · 2026-07-24
- Claude Code and CodeRabbit pair code generation with automated PR review — CodeByPoonam · 2026-07-24
- Tolgee adds an MCP connector for translation keys and machine translation — modelcontextprotocol · 2026-07-24
- Valv open-sources scoped AI database queries with JSON-schema controls — PrestigiousRoof3551 · 2026-07-24