Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 26·5 min read
Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses
TL;DR
Recuris is a new recursive Experiential-Working Memory architecture designed to enhance long-horizon task performance in AI agents. It significantly improves task success rates, achieving state-of-the-art results across multiple benchmarks.
✦ Why It Matters
Consider integrating Recuris into your AI models to enhance their long-horizon task performance today.
Key Takeaways
- Implement Recuris to improve long-horizon task performance.
- Utilize structured evidence collection to identify memory failures.
- Focus skill selection on current task needs rather than historical data.
- Leverage the scalable nature of Recuris for continuous improvement.
How It Works
Recuris combines Working Memory, which tracks current task progress, with Experiential Memory, which stores past experiences. This coupling allows agents to select skills based on immediate needs, rather than relying on a potentially overwhelming history.
The architecture also enables the collection of structured evidence during task execution, which helps pinpoint failures to specific memory components, facilitating targeted updates to Skill Memory.