EverMind (@evermind) header
EverMind (@evermind) profile picture

Memory | Context | Knowledge | Proactivity | RSI

United States
Joined November 2025
Photo by EverMind (@evermind)Photo by EverMind (@evermind)Photo by EverMind (@evermind)Photo by EverMind (@evermind)Photo by EverMind (@evermind)Photo by EverMind (@evermind)Photo by EverMind (@evermind)
EverMind retweeted
I tested @evermind’s EverOS Cloud by giving an AI agent something important to remember: a project decision. For my fictional SpendWise app, I saved: → PostgreSQL as the database → Infrastructure budget under $100/month → A managed database at ~$180/month rejected for exceeding the budget → Reconsider only if traffic or operational needs justify the cost Then I opened a fresh session and searched EverOS for that decision. It retrieved the saved episode and profile, and the next-step recommendation respected those constraints, without resending the original decision. That’s persistent memory in action: save context once, retrieve it later, and use it to inform the next task.
🤝 Paid partnership
130
94
52
122
14,629
EverMind retweeted
EverOS Cloud × Claude Code 🧵 Claude Code built this living blueprint in Three.js: ink shaders on cyanotype paper, stage-by-stage GPU assembly, 17 vehicle kinds. EverOS Cloud turned what it learned into self-evolving skills that any agent, not just Claude Code, can recall.
112
101
65
129
16,321
EverMind retweeted
Every AI I use eventually forgets what I'm building. Chat gets too long, or I'm re-explaining the whole project from scratch Tried @EverMind's EverOS Cloud. Went through the docs and liked how it organizes context into profiles, episodes, and a Memory Bank. Built a small chat app on its APIs. Real use case: I'm building a student resume portal for my college. Students showcase skills and resumes, alumni search by skill, review profiles, and find people for referrals and collabs. Video shows the workflow and the Memory Bank 🚀 Tired of starting from zero? Check it out: everos.evermind.ai
🤝 Paid partnership
149
102
52
133
38,439
And we are live with another rendition of @MiniMax_AI intelligence in the open with @FireworksAI_HQ, @digitalocean, @radixark, @Modular and @evermind
2
1
1
25
10,153
EverMind retweeted
拥有 Harness of harnesses,以及 RSI (递归自我进化)架构的 Raven,国庆节也要休息休息啦! 欢迎大家持续关注 @evermind 最新 harness 架构的更多动态哦!目前我们的技术报告也已经在 arxiv 上发布啦!欢迎大家查看!
模型一直在升级,是不是可以不需要给 「Agents」当项目经理了? 前几天我给大家分享了 Raven 做的游戏、PPT,还有调研和实验案例,在不同程度上说明了现在多 Agent 的能力边界,无论是视觉效果还是功能,都很强大。 现在让 AI 完成一项具体工作,已经不稀奇了。但真正做项目时,我们的精力却经常花在安排顺序、同步信息和检查结果上。 研究报告生成了,要挑出有用的结论交给写代码的 Agent;代码改了,要通知负责测试的 Agent;设计调整了,还要确认其他成员有没有用错版本。 每个工具都节省了一些时间,管理这些工具又成了一份新工作。 这就是 Raven 希望解决的问题:当一项任务需要多个专业 Agent 参与时,怎样减少协调成本,让整个项目持续推进? 多 Agent,其实有不同的组织方式。如 图1 所示。 一种是任务之间相对独立。比如调研几个不同的项目,可以同时执行,最后汇总结果。这类独立并行的协作方式,可以称为 Swarm。 另一种存在明确的前后依赖。比如代码实现完成后才能进行相应的集成测试,游戏资产准备好后才能完成集成。这类按依赖关系组织的方式,通常采用 Graph(图式编排),其中常见的一种结构就是 DAG(有向无环图)。 还有一种强调长期协作。不同成员持续负责研究、开发、审查等工作,保留各自的上下文和经验,在项目中反复参与。这类组织方式可以称为 Agent Team(智能体团队)。 这些方式可以组合使用。选择哪一种,取决于任务有没有依赖、上下文是否需要隔离,以及成员是否需要长期保留经验。 Raven 的任务编排,首先处理其中的依赖问题。 DAG 是工作流系统中常见的结构。每个节点代表一项任务,箭头表示任务之间的依赖。 拿我们之前展示的游戏来说,其中一条资产制作流程包含八个节点: 参考研究、现有资产审查、美术规范、武器制作、Boss 制作、资产集成、构建验证,以及游戏内效果审阅。 参考研究与资产审查可以同时进行。美术规范确定后,武器与 Boss 可以并行制作。集成则必须等到资产准备好。 无环不代表不能修改。一个开发节点内部可以多次测试和修复;当前一轮完成后,也可以根据结果安排新的任务。无环约束的是图中的依赖关系。 节点也不必与 Agent 一一对应。同一个成员可以承担多个任务,没必要为每一步都创建一个新角色。 执行顺序确定以后,还有一个容易被低估的问题:交接。 研究节点显示已完成,开发节点也并不能开始工作。它需要知道采用哪种方法、结论有什么依据、哪些条件还不确定,以及具体该读取哪份文件。 缺少这些信息,后续成员可能重复研究,也可能把一个有前提的结论当成通用事实。 因此,Raven 在调度任务时,也会组织明确引用的上游输出和产物入口。前置任务失败时,依赖它的节点可以被跳过,不相关的独立分支继续执行。 这也涉及上下文管理。研究成员不需要读取所有调试日志,设计成员也不需要了解每一次参数实验。每个成员获得与当前任务相关的信息,同时保留查阅上游依据的入口。 我们希望减少的不只是等待时间,还有重复阅读、重复判断和信息误传。 当然,状态上的“完成”不代表质量已经合格。验收标准仍然需要提前定义。 Harness of Harnesses,是在这个基础上的进一步选择。 Harness 是支撑 Agent 执行任务的运行机制,包括上下文管理、工具调用、规划、检查和错误处理。 不同专业 Agent 已经形成了不同的工作方式。Research 需要检索和证据组织,Code 需要开发与调试,Design 需要视觉制作与交付,Oncall 需要实验运行和持续观察。 Claude Code、Codex 等工具,也各自拥有成熟的执行机制。Raven 希望保留这些专业能力,在更上一层安排成员、任务、依赖和结果交接。成员内部如何执行,则由各自的 Harness 负责。 这就是我们所说的 Harness of Harnesses。 它的价值需要通过完整任务来检验。多一个成员,就会增加一次协调和交接。简单任务交给一个 Agent 完成,往往更合适;复杂任务才需要认真考虑专业分工带来的收益。 但完成一次项目,还不够。 我们更在意的是:这个项目里纠正过的问题,下次能不能少出现一些? 用 AI 久了,很容易遇到这种情况:它能复述你的要求,执行时却还是用了旧方法。 保存一段聊天记录,并不一定会改变后续行为。经验需要进入实际的规划、工具使用和检查过程。 在 Raven 中,EverOS 承接跨会话的上下文与经验,Playbook 保存成员、任务和依赖安排。我们进一步探索的是,让执行中的反馈能够修改 Harness 本身。 这与 RSI,也就是递归自我改进有关:系统利用执行结果和反馈,改进产生这些结果的机制。 这种改进不一定发生在模型参数上,也可以发生在提示、策略代码、工具规则和工作流程中。 例如,一个旅行规划 Agent 经常安排跨区折返。用户指出问题后,后续规划可以增加区域优先的安排方式,并在交付前检查交通与步行负担。反馈如果落实成这样的执行变化,就不需要每次重新提醒。 Raven V0.2.0 中的 Curator 是这方面的实验性实现。它结合任务要求、执行记录和用户反馈生成 Harness 改动,经过声明检查、实际装配与预检后安装;安装失败时,触发恢复流程。 用户提出的要求,由此有机会成为后续任务实际采用的方法。 自我改进最需要验证的,是“改进”这两个字。 修改了提示或代码,只能说明系统发生了变化。当前任务表现更好,其他任务有没有退步?提升是否能重复出现?额外消耗的时间和 Token 是否值得?评价标准能否保持独立? 这些问题不能靠一次漂亮的演示回答。 目前,Curator 已在模拟案例中完成多轮反馈与修改,但跨场景、重复实验和长期效果仍需要继续验证。专业能力评测也不能直接作为自我改进有效的证据。 对 Raven 来说,DAG 负责组织当前任务,专业 Harness 负责执行,记忆和 Playbook 保留可以复用的信息与安排,Curator 则探索如何根据反馈调整工作机制。 这些设计最终应该减少的是人的重复劳动,而不是人的判断权。 目标是什么、哪些行为不能接受、资源有限时如何取舍、什么结果可以交付,仍然需要由人决定。 我们最希望的是一个项目结束时,留下的不只有作品,还有完成这类工作的经验。 下一次面对类似问题,我们能从这些经验继续,而不是重新解释背景、重复同样的纠正。 模型会更新,工具会更换。我们在真实工作中积累的判断,应该能够保存下来,并继续发挥作用。 Raven 0.2.0,带着最先进的理念以及强大的功能,开放给大家!欢迎体验! @evermind
1
6
570
用 Opus 5.5 搓了个产品故事视频! 去年 7 月想做类似的事,受限于大模型的能力无法完成,没想到现在就进步到这个程度了! 如果是去年的我看到这个,直接晕眩瘫软仿佛看到原子弹爆炸。 BTW: 本视频由开源版 Raven v0.2.0 指挥 Claude Code 生成 如果点赞多我会整理成 skill 放出来!@evermind
22
3
1
106
20,397
Two months ago, we, @evermind, launched Raven 🐦‍⬛. The next version is almost here, and I think this game shows where we're headed better than a diagram could. We built this 3D boss fight with Raven. Watch the lighting, the two boss phases, the weapons, and the ending. Please turn the sound on too. I'm genuinely impressed by how it all came together. It took real compute and plenty of iteration. What excites me is Raven coordinating specialized agents on one task graph to make something that feels like one game. That's what we mean by “Harness of Harnesses.” More soon. Raven is open source: github.com/EverMind-AI/Raven
3
1
2
13
13,099
EverMind retweeted
Running several agents gets messy fast. Raven is built to coordinate them. Raven is a pre-alpha multi-agent orchestration repo for builders who want to coordinate specialized agents from one place. It helps you delegate work across research, coding, design, and unattended workflows by bringing built-in and third-party agents into shared workflows. Key features: • Unified orchestration surface – delegate tasks, coordinate execution, and integrate results in one place • Four built-in agents – Raven-Research, Raven-Code, Raven-Design, and Raven-Oncall cover distinct workflow types • Deep research support – Raven-Research produces structured reports with traceable sources for complex questions and technical analysis • Agentic development workflows – Raven-Code supports implementation, debugging, refactoring, data processing, and analysis • Third-party agent connections – connect external agents and coordinate their capabilities in shared workflows It’s open-source under the Apache License 2.0. Link in the reply 👇
6
11
3
39
8,390
EverMind retweeted
The more I look into @evermind, the more I think the real problem they’re trying to solve isn’t intelligence. It’s 𝐜𝐨𝐧𝐭𝐢𝐧𝐮𝐢𝐭𝐲. AI agents can already write code, use tools, call APIs and complete multi-step tasks. But once the session ends, a lot of that context disappears. @evermind is building the infrastructure for agents to carry that experience forward. 🧵🔻
6
4
1
61
2,744
EverMind retweeted
盛大集团带着百万级投资来找项目了,正在做产品的朋友看过来👇 NovaCatapult「超级个体弹射器」正式招募开发者和创业团队,人民币 100 万起投,帮你把想法做成产品,也让好产品有机会被更多人看到。 除了资金,早期创业容易缺的几块支持也是全乎的: 算力:内部集群优先调用,减轻采购 GPU 的压力(做算法的都知道有多少含金量)。 技术:资深工程师帮你评审架构,遇到问题有人分析拆解,提出合理方案。 增长:增长团队协助获客,并有机会对接 @evermind @TankaChat 、@thetawellnessai、@makeplayai、@alayastd、@大圆镜 等生态渠道。 后台:法务、HR、财税等支持,让团队少为产品之外的事情分心。 欢迎独立开发者和小团队,尤其是已经动手、有想法有 Demo,准备全职把项目做下去的朋友。 补充一些小 tips:「超级个体弹射器」是投资孵化,不是无偿补贴;百万方案包含 50 万现金+50 万算力等资源,涉及股权及后续转换条款;核心创始人需要全职到岗,目前以上海、北京为主,具体大家可以到官网查看~ 感兴趣可以直接官网报名,也欢迎 DM 我,发一段项目介绍、目前进展和 Demo 链接。我可以通过 @evermind帮大家做内部推荐,对接后续评审。 期待大家天马行空的作品!
15
19
5
86
23,242
Downloads don’t lie. EverOS OSS: ~400/day, 2.2K/week, ~6K/month. Why: • editable Markdown • local 3-part stack • dual tracks • scoped retrieval • source-backed wiki • offline reflection. Benchmarks remain strong. And the best part is, an even better version of EverOS is coming.
4
1
1
9
2,324
EverMind retweeted
Day 91 — 🥉 3rd Prize + Paid User at @Snowflake × @UofBeta × @evermind Hackathon At the @VoiceOSApp (YC P25) × @a1mobile hackathon, I briefly met Samvrith Bandi. I think we talked for maybe 2 minutes. But I remember thinking: “I want this guy on my team for the next hackathon.” So right there, we found another hackathon, signed up, and just like that, I had my first teammate. When I got to today’s hackathon, I started looking for more people. I always try to go for the max team size. I know quite a few builders now, but I also love finding people I don’t know yet. Especially first-time hackathon people. I have a strangely high success rate with first-time hackathon teammates. There is a reason my shirt says: “First time at a hackathon? Please join me.” Then I met Isha Singh . This time, I actually asked her if she wanted to join the team. Not my usual style, lol. Usually I just say: “You are in my team.” Then I met Rikin Shah, CSM®, FOCP AI on the spot too, and he became part of the team. So that was the team: Sam, Isha, Rikin, and me. Then we started building. Sam did not disappoint at all. We divided the work, and his part was basically: “Build the app.” Not a small task. He single-handedly built a better app than I expected. Not only did he build it, but he also made it work. I already had a feeling he was a GOAT. He proved it. BOY IS KILLING AT 18. ASK ME WHAT I WAS DOING, LEARNING HOW TO USE A MOUSE IN 12TH CLASS. Isha killed every task I gave her. Near the end, we needed to integrate YouTube. None of us had done it before. She figured it out in around 20 minutes. And this was her first hackathon. Not bad for a first one 😄 Ask me — I lost quite a bit before winning one. What stood out about Rikin was his ability to handle pressure. He kept up with the pace. He stayed objective. He did not bring ego into the room. That matters a lot. In hackathons, pressure exposes people quickly. Some people freeze. Some people argue. Some people drift. Rikin stayed locked in and helped put everything together until the last moment. That is rare. We ended up winning 3rd prize. And even better, we got a paid user. That part matters to me because it means the idea was not just interesting to judges. Someone actually saw value in it. Huge thanks to Samvrith Bandi, Isha Singh and Rikin Shah, CSM®, FOCP AI for trusting the process and building hard with me. Hackathons keep reminding me: You don’t always need a team before you arrive. Sometimes you just need to recognize hungry people quickly. Give them trust. Give them ownership. And then run together. @Snowflake @UofBeta @evermind
2
1
1
6
549
EverMind retweeted
Unlock AI self-improvement with Raven – the memory-first agent harness. See link below 👇
1
2
10
1,467
EverMind retweeted
🎉 EverOS 刚刚突破 10,000 ⭐🚀! 感谢每一个 star 的你。为了这个里程碑,我做了一个视频,把 EverOS 的 7个核心能力拆解一下~ EverOS 是一个真正跑在你本地、你能读懂每一行记忆的 AI 记忆系统。非常欢迎大家部署体验并且提建议,我们重视每一个开发者的反馈! 你最想让你的AI记住什么呢? 记得顺手点个 star🌟
9
10
3
73
12,721
今天在杭州参加亚马逊云科技 User Group 社区日 2026 的活动,没想到这么火热,来展台的人非常多,带的礼品都有点不够了。 我们在 2 楼,互动的朋友,我们下次多准备点礼物,把这次缺失的补上。😀
3
1
23
2,680
EverMind retweeted
今天在线下展台介绍 EverOS 和 EverMe,被问到最多的一个问题是: 「Claude 的上下文已经到 1M 了,为什么还需要你们的记忆产品?」 我的理解是,长上下文解决的是「这一轮对话能装下多少东西」;而记忆系统解决的是「一个 AI Agent 能不能跨时间、跨任务、跨身份持续成长」。 1M context 很强,但它不是无限记忆。 当上下文越来越长,系统迟早会触发 compact、summary 或截断。信息会被压缩,细节会被舍弃,很多“当时为什么这么做”的过程也会消失。对一次对话来说这可能够用,但对长期协作来说,这还不是真正可控的记忆。 EverOS / EverMe 想补上的,正是这一层长期记忆能力。 我们做的是跨 Session、跨 Agent 的记忆层:今天在 Claude 里的经验,明天在 Codex 里仍然可以被召回;一个 Agent 做过的 case,另一个 Agent 也可以复用;一个项目里的长期偏好、约束和历史决策,不会因为窗口关闭或上下文压缩而消失。 同时,我们采用双轨记忆机制: 一条是 User Memory,记录人的偏好、习惯、背景、长期目标和约束。 另一条是 Agent Memory,记录 Agent 做过的 case、学到的 skill、踩过的坑和解决方案。 更重要的是,这套记忆不是黑盒。EverOS 会把记忆总结成可读、可编辑、可追踪的 Markdown 文件。你可以打开它,检查它,修改它,删除它,甚至用 Git 管理它。 所以我们不是在和 1M context 比谁更能“塞东西”。 长上下文是更大的工作台,但工作台再大,也会被整理、压缩和清空。 记忆系统是 Agent 的长期经验层:它让 AI 不只是临时读完一堆上下文,而是真的能在一次次任务之后记住、修正、复用,并变得更可靠。
6
3
28
2,817
1) progressive disclosure, i.e. storing content layer by layer with caches and abstractions in between 2) file/markdown + 3) memory cleanup (reflection/dreaming) is becoming the trend. basically a fs but designed for agents to retrieve and manipulate
Introducing Knowledge Wiki and Reflection: two new memory capabilities in EverOS 1.1.0. EverOS can now do more than store what happened. It can organize what agents know, and improve memory between sessions.
1
4
760
Some people say our open-source library EverOS lacks something you can actually play with — that Agent Memory feels too abstract. So I built this: a live demo running as a TUI right in your terminal. Now you can interact with it and visualize the different memory stages inside EverOS in real time. Worth checking out. github.com/EverMind-AI/EverO…
4
1
5
543
EverMind retweeted
Most AI agents still wake up with amnesia. Every new session starts from zero. Your context, decisions, and workflows don't carry over. EverOS is the open-source memory layer built to fix that, and 1.0.0 just shipped. @evermind
113
203
2
250
308,780
EverMind retweeted
Self-evolving memory layer for Claude Code and other agents github.com/EverMind-AI/EverO…
1
12
1
65
5,481