The paper behind the item is Omni-SimpleMem (arXiv 2604.01007), from a UNC-led team: an autonomous research pipeline started from a naive agent-memory baseline and, with no human in the inner loop, executed roughly 50 experiments across two benchmarks over what the newsletter reports as 72 unattended hours, lifting F1 on LoCoMo from 0.117 to 0.598 (+411%) and on Mem-Gallery from 0.254 to 0.797 (+214%) to reach state of the art on both. The authors stress that the biggest gains were bug fixes, architectural changes and prompt rewrites rather than hyperparameter tuning, which is what separates the run from AutoML. In the recursive-self-improvement trajectory it sits between Karpathy’s 650-experiment autoresearch run and Anthropic’s report the next day that Mythos sped internal research up to 400x; the target it optimized, agent memory, is the same one ALMA had agents meta-learn in February.
Development
Three unsupervised days produce a memory system beating every baseline
The author's 'scientist is now a subroutine' moment: an unattended loop is credited with inventing an architecture rather than tuning one, and its largest gains came from bug fixes and structural changes beyond the reach of AutoML.
Claim
UNC researchers let an AI run autonomously for 72 hours, during which it ran fifty experiments and invented a long-context memory system that beats every human-designed baseline.
Reported metric: 72 hours / 50 experiments
Reporting date: (not necessarily the event date)
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to April 7, 2026 (theinnermostloop.substack.com)
- Omni-SimpleMem: Autoresearch-Guided Discovery of Lifelong Multimodal Agent Memory (arxiv.org)
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 →
- Agents meta-learn their own memory architecture
- A model reportedly speeds internal AI research by up to 400x
- Agents review their own session histories overnight
- A model runs thirty-five hours without a human
- A model learns to abandon dead ends better than its researchers
- Agents act as their own data scientists, gaining most by optimizing themselves
- Welcome to April 7, 2026
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.