AFK AI says its product Comma already implements the single-Router, multi-Worker super-agent architecture later outlined by DeepSeek's Harness team, adding persistent memory, continual learning and 24/7 uptime to address the single-session context bottleneck. The system is publicly available with open-sourced code, and the piece frames human attention—capped at roughly 3-5 active sessions—as the real constraint on agent concurrency.
Continual Learning
持续学习
The ability of a model to learn new tasks over time while retaining previously acquired knowledge.
Curated Intelligence
Anthropic published research titled 'The components of agent memory,' splitting Claude agent memory into four layers: session logs, memory notes, knowledge representations and context injection, plus a 'dreaming' mechanism that lets the model review logs, correct errors and distill rules during idle time. On the CL-Bench database QA task, memory-enabled Claude Opus 5.5 and Sonnet 5.5 scored 0.838 and 0.833 on average, well above their no-memory baselines.
Meta's new paper introduces Routed CL, an architecture that uses an LLM router to cluster agent memories by problem type instead of maintaining a single shared history. Experiments show this approach improves scores by 15-30% of the maximum possible gain and reduces estimated API costs by 65-72% compared to shared memory baselines.
Google Research introduces the 'Sleep' framework to address catastrophic forgetting in continual learning for LLMs. By mimicking memory consolidation during sleep, this approach enables models to integrate new facts and enhance reasoning without degrading previously acquired knowledge, overcoming static knowledge cutoff limitations.
Google Research introduces a 'Sleep Phase' framework for Large Language Models (LLMs) to address catastrophic forgetting in continual learning. By simulating biological sleep mechanisms, this approach allows deployed models to offline-consolidate new facts and enhance reasoning capabilities while preserving existing knowledge, overcoming the limitations of static knowledge cutoffs.
Tongyan Lingzhi, a team incubated by BIGAI, proposes 'Recursive Self-Improving Personal AI' to accumulate user-specific knowledge via lightweight parameters and continuous learning. The approach aims to address the limitation of general LLMs in retaining personal or organizational intellectual assets, focusing on building an independently evolving exclusive intelligence layer atop foundation models.
QCon Shanghai previews a talk by Ant Lingbo experts on designing continuous learning systems for embodied AI via full-link data loops. The session highlights that the real world is not a static dataset, requiring edge-cloud collaboration to identify capability gaps and validate improvements. Currently, this is only an agenda preview with no released technical details or open-source assets.
The QCon Shanghai conference will feature a track on 'Data-Driven Embodied AI,' curated by Dr. Huang Yongtao from Ant Lingbo Technology. The session aims to explore scaling embodied intelligence through high-quality real-world data systems, optimized continuous learning loops, and bridging the sim-to-real gap. Currently, the content is limited to agenda previews and topic outlines, with no specific technical releases or open-source projects announced yet.
SCLATE introduces an execution substrate for continual-learning agents, utilizing an open event scheduler and hybrid simulated clock to handle long-horizon multi-session tasks. It enables unmodified agents and benchmarks to share a timeline, compressing month-long scenarios into hours while supporting non-invasive recording and post-training. Experiments show that Qwen3.5-4B post-trained via SCLATE achieves a 16.7-point gain in SWE-bench Verified pass rate with significantly improved memory efficiency.
Google researchers introduce a new 'Nested Learning' paradigm to address catastrophic forgetting in continual learning. Inspired by human neuroplasticity, this multi-level framework enables models to acquire new skills dynamically without sacrificing prior knowledge, offering a novel algorithmic foundation for building autonomous agents with long-term memory.