@rohanpaul_ai: The paper shows that agents reason better over long periods when no past information is thrown away. Keeping every past…
Summary
The paper introduces PRO-LONG, which uses a programmatic, lossless memory system for AI agents, storing all past actions in a structured text log. This approach significantly improves long-horizon reasoning and performance on ARC-AGI-3 games while using fewer tokens than stronger specialized systems.
View Cached Full Text
Cached at: 07/24/26, 05:17 PM
The paper shows that agents reason better over long periods when no past information is thrown away.
Keeping every past action in a searchable log helped coding agents solve long tasks far more reliably.
Long-running agents often lose crucial details because summaries discard information before anyone knows what will matter later.
PRO-LONG instead records every observation, action, result, and short plan in one structured text log.
The agent then uses ordinary coding tools, such as search commands and Python, to recover relevant evidence.
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 to 5.8 times fewer billed tokens.
overall, shows long-term agent memory may work best when storage stays lossless and retrieval becomes programmable.
– arxiv. org/abs/2607.20064
Title: “PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning”
Similar Articles
Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents
This paper introduces a proactive memory agent that runs alongside an action agent to prevent behavioral state decay in long-horizon tasks, achieving significant improvements on Terminal-Bench2.0 and τ^2-Bench. The authors also train Qwen3.5-27B using SFT and GRPO as an early step toward open-weight memory policies.
Proactive Memory for Long-Horizon Agents (16 minute read)
This paper introduces a proactive memory agent that operates alongside a standard action agent to selectively inject memory-grounded reminders during long-horizon tasks, mitigating behavioral state decay. Experiments on Terminal-Bench and τ²-Bench show significant improvements in pass@1, and the approach is demonstrated with both weak and strong action agents.
@yoheinakajima: https://x.com/yoheinakajima/status/2081741659260477666
This thread explores how the brain's dual memory systems (hippocampus and neocortex) offer lessons for building long-running AI agents, arguing that agents need a fast episodic capture and slow consolidation mechanism to avoid catastrophic interference, rather than relying solely on frozen models with temporary scaffolding.
@rohanpaul_ai: A Primer paper about how reasoning models improve after training Shows that better reasoning models depend less on raw …
This primer paper explores how reasoning models improve after training, arguing that effective reasoning data relies more on checkable training evidence than raw data size. It categorizes reasoning data by verification methods and emphasizes preserving messy agent data for learning signals.
@tli104: New paper: "Self-Compacting Language Model Agents" LM agents build up long traces of reasoning and tool calls. As the t…
New paper proposes self-compacting language model agents that can decide when to clean up their own traces of reasoning and tool calls to avoid accumulating mistakes and stale information.