PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning

2026-07-22 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

PRO-LONG is a novel context management framework designed for large language model (LLM) agents tackling long-horizon, exploratory tasks. It addresses the persistent challenge of sustained perception and reasoning in LLMs, which often exhibit limited performance on continual learning benchmarks like ARC-AGI-3. The framework overcomes the tradeoff in existing context management methods—where preserving more information complicates retrieval—by maintaining a complete, structured interaction log and employing coding agents for efficient history search. On the full ARC-AGI-3 public game set, PRO-LONG enhances base coding agent performance by an average of 18.0 percentage points across frontier models. It achieves up to 76.1% pass@1, matching or surpassing specialized harnesses while consuming 4.2-5.8x fewer tokens. Notably, with Fable 5, PRO-LONG reached 97.4% best@2 at a total cost of \$1,750.

Key takeaway

For AI Engineers developing LLM agents for long-horizon or exploratory tasks, you should consider implementing programmatic memory frameworks like PRO-LONG. This approach significantly improves performance on benchmarks such as ARC-AGI-3, offering up to 76.1% pass@1 while drastically reducing token usage by 4.2-5.8x compared to other methods. Adopting a structured interaction log and coding agents for history search can enhance your agent's reasoning capabilities and cost-efficiency.

Key insights

Programmatic memory and structured interaction logs enable LLM agents to efficiently manage context for long-horizon reasoning.

Principles

• Context management significantly impacts LLM agent performance on long-horizon tasks.
• Programmatic memory can resolve the tradeoff between information preservation and retrieval.
• Structured interaction logs facilitate efficient historical data search by coding agents.

Method

Maintain a complete, structured interaction log and leverage coding agents to efficiently search this history for relevant details.

In practice

• Integrate programmatic memory into LLM agent architectures for complex tasks.
• Employ coding agents to navigate and utilize extensive interaction histories.

Topics

• LLM Agents
• Context Management
• Programmatic Memory
• Long-Horizon Reasoning
• ARC-AGI-3 Benchmark
• Fable 5

Code references

• alexisfox7/PRO-LONG

Best for: AI Architect, Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents — LLM agent memory should be actively reconstructed and dynamically adapted to reasoning context, not statically retrieved.
• Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents — The Proactive Memory Agent actively manages and selectively injects relevant memories to combat "behavioral state decay" in long-horizon AI tasks.
• Longer Context Silently Shortens LLM Reasoning — Context length and management significantly impact LLM reasoning efficiency and accuracy.
• LongMINT: Evaluating Memory under Multi-Target Interference in Long-Horizon Agent Systems — Current memory-augmented agents struggle with interference and multi-target reasoning in long, dynamic contexts.
• Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agents — LLM agents can learn unified long-term and short-term memory management via tool-based actions and progressive reinforcement learning.
• TokenMizer: Graph-Structured Session Memory for Long-Horizon LLM Context Management — LLM session history is a structured knowledge artifact, not flat text, enabling efficient context management.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.