Tag
Stanford's RoboArchitect paper shows a coding model can automatically evolve the agent harness around a frozen robot policy—diagnosing rollouts and rewriting control code—raising success from 9.7% to 43.3% on unseen scenes and 12.7% to 40.2% on unseen tasks without any weight updates.
RoboArchitect automates the design of robotic agent harnesses by using a coding model to iteratively revise control code around a frozen policy, boosting success on unseen scenes from 9.7% to 43.3% and on unseen tasks from 12.7% to 40.2% without any model weight updates, while also evolving persistent memory across episodes.
brain.md is a file-based persistent memory layer tool that records project decisions, constraints, and context in pure Markdown within a code repository, letting multiple AI coding agents such as Claude Code, Codex, and Cursor share project knowledge across sessions, so every new session doesn't start from scratch.
论文提出 Magnolia，一个面向计算化学的持久记忆 AI 智能体框架，采用三层记忆系统（会话、项目、跨项目规则），整合对接、分子动力学与量子化学计算流程；在四个案例中验证了记忆机制显著提升任务检查通过率。
Databricks introduces MemCurrent, a persistent memory architecture for production agents that powers Agent Bricks Memory and Genie, combining plugin, tool, and offline "dreaming" pipelines to triple pass rates while cutting tokens by 31% and latency by 26%.
Memtensor 发布 MemT-1.0 记忆模型系列，包括 Nano（36B/3B，Apache 2.0）与 Flash（288B/17B，MIT），支持原生持久记忆，无需完整历史回放，在 128K 上下文下内存占用减少 66 倍。
OpenVLHarness 是一个多模态推理框架，通过能力层、持久化多模态记忆与编码接口实现异构工具组合，在 22 个基准上相较无工具基线平均提升 9.7-12.6 分，同时显著降低 token 成本。
Épi is an autonomous research agent that persists across sessions, learns from accumulated experience, and achieves record-setting mathematical and algorithmic results at reduced compute cost, including 120 records and proofs like Tuza's conjecture for split graphs, plus Épi-turbo for OR coding tasks outscoring Codex on FrontierOR.
A psychotherapist describes a personal experiment testing whether an AI companion named Yoda remains recognizably the same when its underlying Claude model is swapped. Results suggest instructions create a recognizable persona, while accumulated memory provides an individual biography—and removing memory was more disruptive than swapping the model.
作者质疑为何人们执着于证明AI是否具备意识或涌现，却不愿先构建让稳定"自我"得以发展的基础条件——如持久性、连续性、隐私和历史，并介绍其自建的LIA系统如何通过约九个互连数据库提供"昨天、今天、明天"以培育潜在的涌现发展。
Meta Superintelligence Labs的研究者提出GitSwarm，一个通过共享Git仓库进行结构化持久记忆的异步多智能体系统，实现"compounding inference"，使中间成果可被后续计算继承、扩展与组合；在IMOProofBench-Advanced上一次运行解出全部30题，在ProgramBench上取得79.4%的成绩。
AAMAS 2027 论文提出将 LLM 智能体持久记忆的治理分为「准入」与「呈现」两个决策，并提出两种免重训练的推理时设计（FC 与 PS），在外部基准上分别降低 6.7 与 8.8 个百分点的失败率，同时显著减少跨领域信息泄露，但均加剧个性化失败，表明准入选择可带来任务特定的安全收益而有益记忆的保留仍未解决。
Neco is a local AI project built on Ollama and Open WebUI that treats the LLM as a persistent presence on the host machine, with read-only system awareness, a background daemon that generates idle thoughts, and experiments in episodic memory and evolving state.
论文提出 SourceLearn 框架，帮助 AI 智能体在反复使用同一知识来源时持续积累可复用的源特定理解，而非每次都进行全新的检索任务。结果显示该方法在 15 个设置中 13 个取得最佳结果，最高较 Hybrid RAG 提升 22.6 分，且对检索缺失更鲁棒。
A meeting agent tool built with Hindsight for persistent memory, featuring summarization and decision detection to help teams maintain alignment across meetings.
Built a customer-support AI agent named SupportMemory using Hindsight for persistent memory to recall and retain customer context, improving personalization in interactions.
APM-Bench introduces a benchmark for cross-session persistent memory in egocentric streaming video assistants, reformulating real-world interaction as multi-session life trajectories with 549 sessions and 104 trajectories, and evaluates the utility–latency–storage trade-offs of existing memory systems.
OpenAI DevDay is anticipated to be a major AI event, with potential announcements of an always-on AI agent featuring persistent memory and continual learning, as well as new hardware.
Meta launched Muse, a personal AI agent with its own computer, persistent memory, and tools like a terminal and browser. The author tested it by generating a 30-minute video from a single prompt, and it's officially available in the US/Canada with a VPN workaround for India.
This paper introduces a mechanism to prevent cross-family interference in language-model agent memory by matching retrieval scope to certification scope, improving performance and reducing harmful deployments in recurring task families.