Tag
A discussion post highlighting the 'institutional memory problem' for AI agents — learned know-how scattered across vendor-controlled memory features, chat logs, and traces — referencing NeuralTrust's Oct 27 piece and InfoQ's distinction between context, memory, and knowledge, and asking how practitioners store durable, portable agent knowledge.
Beacon, an open-source self-improving memory and context handoff tool for coding agents, is trending #1 on Product Hunt, letting developers resume sessions across 20+ tools like Claude Code, Cursor, and Codex without re-explaining the project.
Waymo for the first time disclosed details about the high-performance compute 'brain' in its robotaxis, while commentators highlight that AI driving systems like Waymo's and Tesla's chips are driving a rapidly growing demand for memory and on-board compute.
SkillMem introduces a compact external latent memory for reusable skills in latent multi-agent reasoning, achieving 38× storage reduction and up to 4.23× inference speedup across benchmarks with Qwen3 models from 4B to 14B.
The author has turned the “Build an Agent from Scratch” tutorial into NamiAgent, an open-source project that serves as both an 11-lesson tutorial with text, images, and code and a locally runnable Agent. It supports reading and writing files, executing commands, MCP integration, long-term memory, and web-based visualization, with only one dependency: the openai package.
Ragionex Memory 已向公众开放公测：它让 AI 代理自动构建并维护一个共享知识网络，配备上下文引擎、MCP 服务器、版本控制与团队共享 Vault，并提供含每月 1 万次请求的免费套餐。
L3G-World 提出了一种紧凑的潜在 3D 高斯记忆（L3G），与视频扩散 Transformer（DiT）交替进行自回归更新，以实现实时、长时程空间一致的交互式无限世界生成，在 9 FPS 下实现了最先进生成质量与显著的效率提升。
一位电子爱好者将从 Facebook Marketplace 购得的 $120 SK hynix DDR5-5600 笔记本内存颗粒移植到 AliExpress 的桌面 PCB 上，打造出价值约 $1000 的 DDR5-7800 桌面内存套装，以应对当前的内存涨价潮。
A user shares excitement about Qwenputer, a novel approach that turns a standard LLM architecture (Qwen3.5-4B) into a file system, a computer, and a new type of memory, hailing it as a potentially revolutionary step beyond the Transformer.
A UC Berkeley study finds agents that periodically drop their history and restart from their best result outperform parallel agents by up to 15%; the piece also cites Microsoft's Fork-and-Flush and Drew Breunig's 'context inertia' to frame the open question of what state should persist across a context reset.
Smart Jelly 是一款 Mac 应用，将日常 AI 工具（听写、会议助手、会前笔记、开放事项与代理追踪）整合到一起，共享同一套用户记忆，避免为多个各不相干的 AI 工具重复付费。
这篇注册研究考察了自改进智能体将运行结果写成教训时，是否会在后续决策中传递运气（随机性）的影响。实验发现，在可靠工具因偶然失败后，LLM撰写的教训会让智能体在后续运行中显著更多地选择较差的工具（11.6% vs 0.4%），而修改记忆写入指令可降低这一危害。
SK Hynix expects demand for its memory chips to outpace supply through 2030, signaling that already high RAM prices may climb further amid AI-driven demand.
Google 等机构提出 StreamVibe 框架，包含 StreamVibe-Bench 基准、StreamVibe-VL 流式主动模型与 StreamVibe-Harness 分层记忆机制，使视觉语言模型能从连续视频观测中推断隐式且动态变化的用户意图，并在响应时机与主动交互上分别提升 10.7% 和 12.3%。
FlexA-Mem 提出一种预算自适应的智能体记忆机制，将历史一次性写入不可变存储，并为每次调用动态渲染紧凑的时间线记忆加相关细节，在 AppWorld、LongMemEval 和 LoCoMo 上显著优于 A-Mem 等现有方法。
SLICE-Mem 提出一种轻量级 LLM Agent 长期记忆框架，通过显著性引导的记忆准入与基于谓词的语义抽取实现免 LLM 调用的在线记忆构建，在保持接近 MemOS 基线精度的同时将 token 消耗降低 92.4%–99.5%。
论文提出 MemCryst 框架，通过记忆与技能的双向学习实现冻结语言模型下的终身上下文适应，并在 ALFWorld 等十个设置中的七个对比中取得领先结果。
提出 Self-Direction 机制，让智能体在任务失败后自我反思并将教训蒸馏进轻量级 LoRA 适配器，构建可持久复用的教训库，在 GAIA、Claw-Eval 和 Terminal-Bench 2.1 上将成功率提升 8.9 至 17.0 个百分点，并显著减少重复犯错。
论文提出 DraMaze，一个多智能体管线，利用 trope memory 引导的进化树搜索与 Backward Narrative Dependency Tree 生成长篇微短剧剧本，并发布双语基准 MiDramaBench，实验显示因果模式重复减少 32.1%，逻辑得分超越基线 8.5 分。