EverMind Raven

• 1. overview (635 chars)

  EverMind Raven (Chinese name "Raven") is a deep self-evolving agent framework (Agent Harness) released on July 8, 2026 by EverMind, an AI technology team incubated by Shanda Group. It is built on EverMind's self-developed memory operating system EverOS. Through a two-way memory mechanism and the evolutionary ability to rewrite its own code, it attempts to push AI from a "passive response tool" to an "actively evolving digital life." It is officially positioned as "The Self-Improving Agent Harness" - an intelligent agent base that is continuously improved around memory, tools, skills, code execution and strategies. Raven is currently open source, supports macOS and Windows, is hosted on GitHub, and is released under the Apache-2.0 license.

• 2. Product background (1017 chars)

  For a long time, large models have had a basic pain point—AI’s “amnesia.” When the user closes the conversation window, no matter how in-depth the previous communication was, the AI ​​will instantly clear everything and return to the initial state. This technical limitation, called "stateless", has become a bottleneck for AI to evolve into a real intelligent agent. In the past two years, the industry has generally adopted RAG (retrieval augmentation generation) or continuously expanding the context window to cope, but the EverMind team believes that these solutions are still essentially "advanced bookmark systems" - "real memory is not retrieval, but internalization."   The judgment proposed by EverMind is: Memory is the dividing line between "tools" and "intelligent agents". An entity without memory cannot exhibit consistency, cannot learn from experience, and cannot evolve. Based on this, EverMind has planned four stages of digital life evolution: L1 role-based command body (most current shell AI), L2 memory enhancement body, L3 self-evolving body, and L4 fully autonomous digital life. The release of Raven represents a substantial step forward for EverMind towards the L3 stage.

• 3. core architecture (1140 chars)

  Raven relies on the "four-layer bionic architecture" underlying EverOS and operates collaboratively through the proxy layer, memory layer, index layer and interface layer. The original dialogue flow is divided into independent memory units, and then a "memory scene" is formed through a clustering algorithm, and finally an in-depth portrait including the user's identity, preferences, skills and work goals is constructed.   At the code level, Raven's GitHub repository shows that its runtime is organized around a backbone called Spine: each round of interaction flows through a single entrance and a single exit of submit → lanes (responsible for sorting and cancellation) → emit (output). Each feature engine accesses the Agent Loop through explicit handoff rather than mutual import. Feature engines include: Context Engine (context assembly and Curator path), Memory Engine (EverOS memory, local skills, SkillForge), Proactive Engine (Sentinel scheduling and reminder strategy), TokenWise (usage tracking and cache routing), and Eval Engine (task judgment and coordination). The repository is currently in pre-alpha, but the core product interfaces (native TUI + CLI, Spine runtime, base Agent Loop, Context Engine, Sentinel initiative, TokenWise, SkillForge) are available and the Eval Engine is still partially implemented.

• 4. Two-way memory mechanism (1239 chars)

  Raven's memory is bidirectional and dynamic. It is based on the in-depth profiling capabilities of EverOS and internalizes the key information in each interaction into a cognitive model of the user - not "looking through your files", but "continuously updating its understanding of you with your experiences." It will remember that you "drink coffee without sugar" and "write an outline first" without your knowledge, and these preferences will become part of its cognition.   What's more, this memory works both ways: Raven not only remembers you, but also draws reflections and directions for improvement on her abilities from every interaction with you. If you correct one of its mistakes, it will not only remember the correct answer, but it will also reflect on why it made the mistake and improve it in the next interaction. EverOS 1.1.0 introduces a three-layer memory classification method - user memory (defining "person"), agent memory (defining "agent") and world knowledge (Knowledge Wiki, defining "world"), and is equipped with a Reflection mechanism inspired by human meditation, allowing agents to automatically integrate and consolidate intelligence during idle periods.   Test data shows that with the support of EverOS, Raven can achieve an accuracy that exceeds Full Context with about 1/10 of the token consumption of traditional solutions. The official benchmark LoCoMo reaches 93.05%, LongMemEval reaches 83.00%, and HaluMem reaches 93.04%.

• 5. Self-evolution ability (961 chars)

  Raven's core and most topical ability is "rewriting its own code." It can not only rewrite its own skills, runtime logic and strategies, but can even dynamically fine-tune model weights through the user-side personalized small model EverBrain. This means that Raven can extract best practices from successful tasks, identify room for improvement from failures, and write insights directly into code to achieve internalization of experience and iteration of capabilities.   Raven has approximately 100,000 in-depth evaluated skills built-in, covering a wide range of scenarios from daily productivity to professional vertical fields. These skills are not static: in actual use, Raven will continue to evaluate which skills are useful and which ones are ineffective, and then actively eliminate, strengthen or combine new skills. This "skill memory" is non-parametric and self-evolving - EverOS records the trajectory of the agent as Cases and refines recurring patterns into reusable Skills, allowing the agent to learn from previous experiences and continue to improve over time, rather than starting from scratch for every task.

• 6. Features and deployment (1083 chars)

  Based on the design concept of "memory first", Raven provides native terminal TUI and CLI interfaces, and has built-in gateways to connect to messaging platforms such as Telegram, Discord, Slack, Matrix, WhatsApp, and Enterprise WeChat. It has a built-in tool manager that can automatically install, update or remove external tools, and supports multi-model switching (Anthropic, OpenAI, Claude, etc.).   Installation uses a one-line command: macOS/Linux execute curl -fsSL https://raven.evermind.ai/install.sh | bash, Windows use PowerShell irm https://raven.evermind.ai/install.ps1 | iex, and then raven start to start. Developers can encapsulate successful workflows into reusable Agent Templates or "digital workers", and use Raven Builder to define exclusive Agents for any scenario and share them with one click.   As an open platform, Raven's memory module, initiative engine, and tool routing are completely independent, allowing developers to freely replace any component without touching the core framework. Its highly decoupled modular architecture forms a positive flywheel: the creation of more agents brings more usage data, and the data feeds back EverOS to make the memory system more accurate, thereby promoting the faster evolution of new agents.

• 7. Application scenarios (1133 chars)

  Multi-agent system: Orchestrating multiple AI agents requires complex coordination and task allocation. Raven's Group Memory and Memory Bank allow cross-platform agents to share knowledge and team context.   Personalized AI companions: From personal assistants to healing chatbots, companion AI requires true long-term memory to bring emotional intelligence and behavioral consistency.   Enterprise knowledge base: Build an organizational memory that evolves with the team. AI learns from internal discussions, documents and decisions, becoming the enterprise's institutional knowledge layer.   Customer support intelligence: Remember historical issues, resolution records, customer preferences and communication styles to make support "warm and continuous".   Wearable hardware: Process fragmented interactions, visual cues and voice commands into structured patterns, allowing the hardware to proactively predict needs.   Single-person productivity: Business users can send voice messages on WeChat to "organize international meeting minutes, refine three core tasks, enter them into the project system and notify the person in charge by email." Raven can complete the process in one stop. EverMind envisions a future where "single person + intelligent agent" can handle the workload of dozens of people in the past.

• 8. Academic and Ecological (994 chars)

  EverMind’s academic team has simultaneously built a theoretical moat, and has recently published several top-level papers: MSA (Memory Sparse Attention) is an end-to-end trainable sparse attention mechanism, extending the context to 100 million Tokens, with performance attenuation of less than 9%, and reached the top of HuggingFace Daily Papers on the day of release; HyperMem hypergraph hierarchical memory architecture reached 92.73% SOTA in LoCoMo, and won ACL 2026 Oral; EverMemOS self-organizing memory operating system was selected as the main conference of ACL 2026; multi-party collaborative dialogue long-term memory evaluation filled the industry gap and won the KDD 2026 Oral.   GitHub Stars exceeded 10,000 within about a month after EverOS was open sourced, and the growth rate exceeded the level of about 7,000 in the first seven months of mem0. EverMind's full-stack ecosystem covers EverBrain (personalized memory model), EverOS (memory operating system and Agent infrastructure), Raven (self-evolving Agent Harness), EverMe (personal memory center), and practices "technology for good" with the ReUnite public welfare talent search platform.

• 9. Evaluation and Outlook (993 chars)

  Raven represents a substantial leap from L2 (memory enhancement) to L3 (self-evolution) of self-evolving agents. Its concept of "memory is not an archive but a neural network" and "not finding but realizing", as well as the indicator of achieving super-full context accuracy with 1/10 Token consumption, are indeed attractive at the paradigm level. Open source first, Apache-2.0 protocol and one-line command deployment significantly lower the threshold for enterprise and personal use, which is conducive to building an underlying developer ecological moat similar to Android/Windows.   But objectively speaking, Raven is still in pre-alpha, the API may change at any time, and it is not suitable to be understood by a mature production system. Its radical ability to "rewrite its own code" not only brings iterative efficiency, but also raises engineering and governance issues such as interpretability, safety boundaries, and risk of loss of control. For developers, Raven is currently more like an early but ambitious Harness sample, which is worth tracking, but when implemented in key businesses, stability and monitorability still need to be carefully evaluated.

No in-depth report yet

• 

  KellyBarnesX

  —

  EverBrain's small user-side model can dynamically fine-tune the weights. This is very important, which means that the memory truly feeds the model itself, instead of just patching the prompt words.

---

• 

  BreadBudNielsen

  —

  "Memory is not retrieval but internalization" is right to the point. Traditional RAG is just an advanced bookmark, but Raven's two-way memory is like a human being.

---

• 

  MChavezII632

  —

  I connected Raven to Discord and Slack. Group memory allows several channels to share context, which is very convenient for making a prototype of a team knowledge base.

---

• 

  SHernandezQ

  —

  As a developer, the most surprising thing is the decoupling of the architecture. The memory, tool routing, and initiative engine can be changed independently without being tied to the core framework.

---

• 

  薛明强

  —

  EverOS has exceeded 10,000 Stars in one month since it was open sourced. The growth rate is indeed rapid, and the community atmosphere has also improved. Following the pace of mem0, it is obviously faster.

---

• 

  RasmusChristiansen

  —

  The token consumption is only one-tenth of the full context and the accuracy is super high. The cost of processing long documents is visibly reduced, which is very practical for small teams.

---

• 

  Sandra.Kim_Pro

  —

  It can really condense a successful workflow into a reusable Agent Template. I used it to organize three days of meeting minutes, and then reused the template directly, and the efficiency took off immediately.

---

• 

  jEFFREYrOGERS

  —

  After playing with Raven for a night, the memory is really preserved across sessions. When we talk the next day, it still remembers the preferences I mentioned yesterday, which is much better than a bunch of shell assistants.

---

• 

  ZHillIII

  —

  Objectively speaking, the story of L3's self-evolution is beautifully told, but the security and loss of control risks caused by "rewriting its own code" are not discussed enough. I suggest that before you really implement the key business, you should first stress test it as a black box: deliberately feed error feedback to see if it will solidify the error experience into skills, and then check whether the skills generated by SkillForge have any unauthorized calls. Raven is now more of an ambitious early prototype that's worth tracking, but it's too early to use it as a mature platform.

---

• 

  Megan_Campbell_2021

  —

  I tried to use Raven Builder to define a dedicated Agent for customer service scenarios and share it to EverMe with one click. The whole process was simpler than imagined, and even non-technical colleagues can get started. Our small team used it as a pre-sales Q&A assistant. After feeding historical work orders into it, its answers were significantly more coherent than the previous version that used fixed prompt words. Customers also reported that "this time I finally remembered what I asked last time." However, it occasionally gets stuck writing to the Memory Bank during peak periods. It is estimated that EverOS needs to further optimize throughput.

---

• 

  ASICant433

  —

  The number of 100,000 built-in skills is too exaggerated. In actual testing, it can be used in common scenarios out of the box. However, I later carefully looked through the skill library and found that many skills in vertical fields are only at the scaffolding level. If you really want to go into production, you will have to adjust it again. You cannot fully believe this number. It is recommended that the official make the "evaluated" standards transparent, otherwise novices will easily mistakenly think that it is all-powerful right out of the box.

---

• 

  GregoryGray_2021

  —

  These benchmarks of LoCoMo 93% and LongMemEval 83% are really reproducible. I ran the official open source evaluation script and the data is basically consistent. This is better than many people who can only shout slogans.

---

• 

  iDylanLewis_dev

  —

  "It can rewrite its own code" sounds cool, but it also makes people a little panic. I used it to run a data cleaning task. It did change the order of tool calls to increase the success rate, but I couldn't see what it changed at all. Who will take the blame if it changes its strategy wrongly? I hope the official government will come up with interpretability and rollback mechanisms soon, otherwise I will not dare to take on key business with this kind of black box self-evolution.

---

• 

  DVasquez_2021

  —

  It's still pre-alpha, and the API changes at will. The script I wrote last week won't run this week. Use it with caution in the production environment. It's okay to use it as a toy or experimental platform. My own approach is to put all configurations and Agent Templates into the repository, and diff them before each upgrade before deciding whether to follow them. This way at least I won't be crashed by a breaking change in the middle of the night.

---

• 

  MoniqueRamos

  —

  It is installed with just one line of commands. After running curl install.sh and raven start, you can directly enter the TUI, which is very friendly to terminal users.

No reviews yet