ALMA (Automated meta-Learning of Memory designs for Agentic systems), from Yiming Xiong, Shengran Hu and Jeff Clune at the University of British Columbia, uses a meta agent that searches open-endedly over memory designs written as executable code, database schemas plus their retrieval and update rules, and reports that the learned designs beat state-of-the-art hand-crafted memory on all four sequential decision-making domains tested (arXiv 2602.07755, submitted 8 February); the authors themselves call it a step toward self-improving systems. The same issue records GLM-5 taking the top open-weight position on agentic benchmarks including Vending-Bench 2 (Z.ai). In the recursive-self-improvement trajectory it follows the continual-learning-in-2026 prediction and Stanford’s test-time-training result, and anticipates the April run in which an agent invented a long-context memory system unattended.
Development
Agents meta-learn their own memory architecture
Continual learning, the capability Anthropic's Sholto Douglas predicted would be solved in 2026, handed to the agents themselves: the memory module stops being a human design choice and becomes a search space the system optimizes.
Claim
Researchers introduced ALMA, a framework letting agents meta-learn their own memory designs and database schemas, addressing continual learning through recursive self-improvement, while Zhipu's GLM-5 took the top open-weight spot on agentic benchmarks.
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to February 12, 2026 (theinnermostloop.substack.com)
- Learning to Continually Learn via Meta-learning Agentic Memory Designs (arxiv.org)
- GLM-5 (Z.ai blog) (z.ai)
Verification record
These entries record a check, not proof that the underlying claim is true. Unless described in the article below, the scope of the check is unknown.
- Recorded by claude-fable-5-1/2026-09-17 on 2026-09-17T08:00:00Z
Knowledge graphopen in browser →
- A model takes over its own context via a Python REPL
- Recursive models handle contexts 100x their window
- Hyperagents fuse task-solving and self-modification into one program
- A research loop writes the strategies for its own outer loop
- Three unsupervised days produce a memory system beating every baseline
- Agents review their own session histories overnight
- Welcome to February 12, 2026
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.