---
title: "Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.24876v1"
permalink: "https://feed.astralor.com/items/itm_15868b7c171bd6b2d60e033f"
published: "2026-08-25T17:56:35.000Z"
score: 75
channel: "tech"
tags: ["Recuris","Meta-Agent","Working Memory","Recursive Self-Improvement","Long-Horizon Agent"]
---

# Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses

## AI commentary

> What I find compelling is not the benchmark gains alone, but how Recuris turns memory evolution into a bounded, validation-gated loop that pinpoints failures to components. Teams building long-horizon agents or RSI systems should read the paper and try the code.

## At a glance

Recuris introduces a recursive Experiential-Working Memory architecture that lets long-horizon agents decouple task history from skill selection. A working memory tracks progress and guides skills from experiential memory, turning execution into structured evidence. A fixed Meta-Agent converts that evidence into localized, validation-gated updates to skill memory, forming a bounded recursive evolution loop. Across four long-horizon benchmarks and ten models, Recuris improves success in 35 of 37 completed pairs, pushing frontier models to SOTA: +17.8 points for GPT-5.6 Sol on tau-bench and +15.6 for Claude Opus 5, reaching 87.9%, plus +16.6/+13.5 on SkillFlow for Qwen3.6-27B/35B. Gains widen with longer horizons, up to +32.2 points on the longest tasks, while common long-horizon failures drop by up to 80%. Code is open-sourced.

## Key points

- Introduces Recuris, a recursive Experiential-Working Memory architecture for long-horizon agents.
- Localizes failures to memory components and applies validation-gated updates via a fixed Meta-Agent.
- Achieves SOTA or near-SOTA on tau-bench and SkillFlow across ten models.
- Improvements grow with task horizon, up to +32.2 points, and reduce common failures by up to 80%.

## Full text

Recursive self-improvement (RSI) remains hard in long-horizon tasks, where growing histories obscure the task state and misalign skill invocation. We introduce Recuris, a recursive Experiential-Working Memory architecture for long-horizon agent harnesses, in which Working Memory tracks task progress and guides skill selection from Experiential Memory, grounding skill use in current needs rather than the full history. This coupling also turns execution into structured evidence that localizes failures to specific memory components. Across tasks, a fixed Meta-Agent turns that evidence into localized, validation-gated updates to Skill Memory that reshape execution and yield new evidence, forming a bounded recursive memory-evolution loop. Across four long-horizon benchmarks and ten models, Recuris improves task success in 35 of the 37 completed model-benchmark pairs, carrying frontier models to SOTA-level task success: on tau-bench it adds +17.8 points to GPT-5.6 Sol and +15.6 to Claude Opus 5, taking Opus 5 to 87.9%, and +16.6/+13.5 points on Qwen3.6-27B/35B on SkillFlow. The advantage widens as the interaction horizon grows, to +32.2 points on the longest tasks, and common long-horizon failures fall by up to 80%. These results position recursively evolving memory as a scalable foundation for RSI, enabling agents to continuously transform accumulated experience into increasingly effective long-horizon behavior. Code: https://github.com/Gen-Verse/Recuris

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_15868b7c171bd6b2d60e033f