DeepAgent folds agent memory while it discovers tools on the fly

Researchers at Renmin University of China posted DeepAgent on 27 October 2026. It is a preprint on arXiv, not yet peer reviewed. The agent finds tools on its own and folds long chat history into structured memory.

The short version

DeepAgent runs thinking, tool search, and actions in one long reasoning stream. Renmin University NLPIR team adds memory folding and ToolPO training so the agent can handle huge tool lists without losing the plot.

What the researchers did

Xiaoxi Li, Wenxiang Jiao, Ji-Rong Wen, Zhicheng Dou, and colleagues at RUC NLPIR wrote the paper. They posted it on arXiv as 2510.21618 on 27 October 2026. Code and datasets sit on GitHub under RUC-NLPIR/DeepAgent.

Most older agents loop in fixed steps like reason, act, then read feedback. DeepAgent keeps one continuous thought. It can search a large tool catalog when a step needs a new skill.

When traces grow too long, the model triggers Self driven Memory Folding. It compresses history into three stores. Episodic memory logs key events. Working memory holds the current sub goal. Tool memory stores what worked with each API.

ToolPO is reward based training tuned for tools. An LLM based simulator stands in for real APIs during training. Fine grained credit goes to tokens that call tools correctly.

Photo: Tima Miroshnichenko / Pexels

What they found

The team tests on broad tool test suites such as ToolBench, API Bank, TMDB, Spotify, and ToolHop. Tool counts range from dozens to more than sixteen thousand RapidAPI entries in demos.

They also test embodied and research style tasks. Those include ALFWorld, WebShop, GAIA, and Humanity's Last Exam style suites. Public README figures show DeepAgent ahead of baselines on those grouped tests.

Memory folding lets the agent pause, summarize progress, and restart without replaying every raw log line. That design targets long jobs where early wrong guesses would otherwise waste context.

Photo: Mikhail Nilov / Pexels

Why it matters for you

Product teams building copilots should watch agents that pick tools dynamically. Fixed plugin lists break when vendors add APIs weekly. Security leads should ask how tool discovery interacts with permission models. A model that can call new endpoints needs tighter policy checks than a fixed three tool bot.

Investors comparing agent startups should check whether demos use simulated APIs or live ones. DeepAgent training relies on simulators for stability.

What to keep in mind

This is a preprint, not yet peer reviewed. Test suite gains may shrink on messy enterprise tools with auth and rate limits.

Large tool search adds latency and cost. Memory folding helps context limits but still uses compute.

WWW 2026 acceptance is noted in the repo news feed. That is a venue signal, not a guarantee for your private data.

School students curious about tech can follow the GitHub links. You do not need a lab badge to read code and watch issue threads. That is a free way to learn how research turns into software.

Small clinics, farms, and shops should ignore hype posts. Ask vendors whether their product cites a peer reviewed paper or only a blog recap. A clear source link is a sign of respect for your time.

Teams that ship fast should still read the full PDF tables. Headline wins on clean test suites often fade once real data adds noise, broken tools, and slow networks.

Readers at home feel this work through safer apps, cheaper goods, or better research tools over time. None of these papers alone change your phone tonight. They show where labs spend effort this month.

If you train models, log which paper version you used. Authors may post fixes after peer review. A date stamp in your internal wiki prevents stale citations in slide decks.

Key takeaways

• DeepAgent merges tool discovery, action, and reasoning in one stream.
• Memory folding splits history into episodic, working, and tool stores.
• ToolPO trains tool calls with simulated APIs and token level credit.
• Posted 27 October 2026 on arXiv from RUC NLPIR with open code.
• Agent builders should test live API auth before copying test suite scores.

Sources

• Primary — arXiv, DeepAgent: A General Reasoning Agent with Scalable Toolsets (27 October 2026). Defines memory folding, ToolPO, and test suite claims.
• Secondary — GitHub, RUC-NLPIR/DeepAgent (2026). Code, datasets, and reproduction commands.