PRO-LONG gives LLM agents searchable programmatic memory and lifts ARC-AGI-3 by 18 points
dair_ai · x · 2026-07-24
PRO-LONG proposes a minimal context-management framework for LLM agents that stores the full interaction log as structured, searchable memory instead of compressing it away.
Core idea
- Keep a complete record of everything the agent has seen
- Query that history on demand with coding-agent tooling
- Treat memory as a searchable database to avoid the usual tradeoff between saving too much and retrieving the right detail
Reported results
- On the full ARC-AGI-3 public game set, it improves over a base coding agent by 18.0 points on average across frontier models
- It matches or exceeds specialized harnesses at up to 76.1% pass@1 while using 4.2–5.8× fewer tokens
- With Fable 5, it reaches 97.4% best@2 at a total cost of $1,750
More from coding & agent
- Luke Wroblewski says agentic software should summarize work, not expose every step — LukeW · 2026-07-24
- LukeW's Deep Dive: Agentic UI Should Collapse Process and Focus on Outcomes — LukeW · 2026-07-24
- Stanford's re:AGENT Build Weekend Focuses on AI for Science Agents — james_y_zou · 2026-07-24
- Databricks Genie runs as an MCP server inside LangGraph, then ships to Azure ML — Cautious-Meringue554 · 2026-07-24
- PenguinHarness open-sources a TypeScript agent builder that claims $0.02 RAG apps — RepulsiveBad8681 · 2026-07-24
- MCPForge scores public MCP servers on security, compliance, and quality — Competitive_Ad_1228 · 2026-07-24