🧠 An agent’s long-term memory can be programmed and improved by the agent itself.
Over a year ago, we began exploring whether the layering principles behind computer networks’ OSI model could inspire a memory framework for agents, with clearly defined layers and shared interface

Arxiv：arxiv.org/pdf/2609.39765

Across four benchmarks, the full Qwen3.5-27B configuration achieves a 10.1% relative improvement in average accuracy over the strongest baseline, MemPro, while using approximately 71% fewer context tokens and delivering 2.1× faster inference.

One key design: higher-level memories serve as indexes into the original evidence. At query time, the system navigates the hierarchy as needed, stopping once sufficient evidence is found while preserving traceability to the underlying records.

Through program evolution, it rewrites memory-management code to improve what to store, how to retrieve it, and when to dig deeper, using evaluation to select effective changes.

Looking up a detail, connecting evidence across conversations, and reusing a skill each place different demands on how information is organized and retrieved. MemCodex organizes raw records, summaries, structured knowledge, and reusable skills into a memory hierarchy

Our new work, MemCodex, takes an initial step toward this vision—and goes further: enabling agents to design their own memory systems.
Different tasks call for different ways of remembering.

The “quantum brain” hypothesis sparked our curiosity: what could quantum computing bring to large language models?

After more than a year of research, we’re sharing HyperQ, which integrates trainable quantum circuits into a 1.1-billion-parameter diffusion language model.

📃arXiv：arxiv.org/abs/2609.24657
🤗Huggingface：huggingface.co/papers/2609.24…

These results do not establish a quantum computational advantage. They give us a concrete starting point for exploring larger quantum foundation models—and understanding what quantum computation can contribute to learning and generation.

With support from IVADO–PINQ², we also ran the 16–64-qubit circuits on IBM’s 156-qubit processor in Québec, retaining most of the model’s performance under real hardware noise on selected benchmark subsets.

Using just one-tenth of the fine-tuning data used by the classical baselines, the 64-qubit version outperformed LoRA on the same backbone by 3.67 percentage points, averaged across six benchmarks.

HyperQ is a lightweight circuit hypernetwork generates circuit parameters dynamically from each token’s context. Across the three circuit sizes we tested—16, 32 and 64 qubits—performance improved as the circuits grew wider.

Introducing ComBodied Agents — our new framework that puts human agency and well-being at the core of Agentic AI, not just task completion.

Proud to share our work. 🧵👇

📄 arXiv: arxiv.org/abs/2608.10915
🤗 HF: huggingface.co/papers/2608.10…

I gave a talk on ``Beyond Capability: Toward agents we can trust, afford, and live with'' for Duke-NUS-Chen Institute Boot Camp for AI & Applications. Slides available on my homepage.

Agents are moving from tool use to paid work.
But paid work needs more than wallets.
It needs contracts, evidence, settlement, and reputation.

That’s Foundation Protocol:
the accountability layer for the agent economy.

Watch the feature overview ↓

check our blog here: foundation-protocol.pages.dev

🤗：huggingface.co/papers/2605.23…
📄：arxiv.org/abs/2605.23218
💻Founda Protocol：github.com/FoundationAgen…
💻Demo：github.com/FoundationAgen…

🫱 Introducing 𝐍𝐞𝐮𝐫𝐚𝐥 𝐂𝐨𝐦𝐩𝐮𝐭𝐞𝐫s:

𝐰𝐡𝐚𝐭 𝐢𝐟 𝐀𝐈 𝐝𝐨𝐞𝐬 𝐧𝐨𝐭 𝐣𝐮𝐬𝐭 𝐮𝐬𝐞 𝐜𝐨𝐦𝐩𝐮𝐭𝐞𝐫𝐬 𝐛𝐞𝐭𝐭𝐞𝐫, 𝐛𝐮𝐭 𝐛𝐞𝐠𝐢𝐧𝐬 𝐭𝐨 𝐛𝐞𝐜𝐨𝐦𝐞 𝐭𝐡𝐞 𝐫𝐮𝐧𝐧𝐢𝐧𝐠 𝐜𝐨𝐦𝐩𝐮𝐭𝐞𝐫 𝐢𝐭𝐬𝐞𝐥𝐟?

Beyond today's conventional computers, agents, and mputers (NCs) are new frontiers where computation, memory, and I/O move into a learned runtime state.

We ask: whether parts of runtime can move inward into the learning system itself. This is our first step toward the Completely Neural Computer (CNC): a general-purpose neural computer with stable execution, explicit reprogramming, and durable capability reuse.

Work done with Mingchen Zhuge (@MingchenZhuge), Changsheng Zhao, Haozhe Liu (@HaoZhe65347 ), Zijian Zhou (@ZijianZhou524 ), Shuming Liu (@shuming96 ), Wenyi Wang (@Wenyi_AI_Wang ), Ernie Chang (@erniecyc ), Gael Le Lan, Junjie Fei, Wenxuan Zhang, Zhipeng Cai (@cai_zhipeng ), Zechun Liu (@zechunliu ), Yunyang Xiong (@YoungXiong1 ), Yining Yang, Yuandong Tian (@tydsh ), Yangyang Shi, Vikas Chandra (@vikasc), Juergen Schmidhuber (@SchmidhuberAI)

Given a fixed budget: scale out with more agents, or scale up a single one?
We derive a sharp criterion comparing the "organization exponent" to the single-agent scaling exponent to determine when multi-agent expansion wins and when it doesn't.

📑 arXiv: arxiv.org/pdf/2601.17311