PRO-LONG
par alexisfox7
Programmatic memory for long-horizon LLM agents: the harness appends everything to one log, and the agent searches it with code. 97.4% on ARC-AGI-3 (arXiv:2607.20064)
Depuis le README
PRO-LONG Programmatic memory for long-horizon agents! PRO-LONG is a minimal memory addition for LLM agents on long-horizon tasks that keeps a complete interaction log outside the accessed context window; the agent searches this log programmatically when it needs earlier history. Our harness achieves 97.4% best@2 on ARC-AGI-3 at a total cost of $1,750. Coding Agent Memory · Research · Paper Figure 1 from our paper. Coding Agent Memory Tool We find that PRO-LONG works beyond ARC-AGI-3 and NetHack, and can aid standard coding agent usage. Add local, persistent, and searchable memory (i.e.,…
Extrait du README du projet sur GitHub.
Skills
Un projet open source — explore le code et auto-héberge-le depuis GitHub.