alexisfox7/PRO-LONG

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)

What it solves

PRO-LONG addresses the challenge of long-horizon reasoning in LLM agents, specifically for complex tasks like those found in the ARC-AGI-3 game set. It prevents agents from losing track of progress or forgetting critical observations over long sequences of actions by providing a persistent, programmatic memory.

How it works

Instead of relying solely on the model's context window, PRO-LONG uses a minimal memory addition: a structured log.txt file that records every observation, action, and outcome. The agent can then retrieve and reason over this history programmatically using tools like grep and Python, rather than having the entire history shoved into the prompt. This approach uses a very short system prompt (approximately 30 lines) and avoids complex sub-agents or specialized retrieval systems.

Who it’s for

It is designed for developers and researchers working on LLM agents capable of solving long-horizon reasoning tasks and those specifically targeting the ARC-AGI-3 benchmark.

Highlights

  • Significant Performance Gains: Improves over standard coding agents by 18 percentage points on the ARC-AGI-3 public game set.
  • Token Efficiency: Matches or exceeds specialized harnesses while using 4.2–5.8x fewer billed tokens.
  • High Accuracy: Reached 97.4% best@2 with Fable 5.
  • Flexible Backends: Supports both OpenAI Codex and Claude Code CLI backends via Docker containers.

Related

  • Project
    MemoriLabs/Memori

    A persistent, structured memory layer for AI agents and LLMs that automatically captures and recalls user preferences and agent actions to reduce token costs and maintain context across sessions.

  • Project
    zjunlp/LightMem

    A lightweight memory management framework for LLMs and AI agents that provides efficient storage, retrieval, and update mechanisms for long-term memory capabilities.

  • Project
    LycheeMem/LycheeMem

    A lightweight long-term memory framework for LLM agents that uses structured semantic, procedural, and visual memory stores to maintain persistent context across sessions.

  • Project
    RichmondAlake/memorizz

    A Python framework for building memory-augmented AI agents featuring multi-layered persistent memory, continual learning, and a control plane for orchestrating multiple agent workers.

  • Project
    aiming-lab/SimpleMem

    A unified memory stack for LLM agents that provides semantically lossless compression for text and multimodal data, featuring a self-evolving retrieval system to maximize recall while minimizing token usage.