Another paper from us, this time on video token compression. Video VLMs have a token problem. A single video can produce tens of thousands of visual tokens, making prefill expensive and KV caches large. But aggressively pruning them risks deleting the details needed for temporal reasoning. CRAFT takes a different approach. It separates two decisions: Which tokens should merge? Global similarity decides. How should they merge? Position-aware weights and a channel-wise gate learn how to preserve the information. The process is recursive and query-agnostic, so the video can be compressed once and reused across multiple questions. At roughly 8× compression, CRAFT retains 96.8% of the backbone’s average accuracy while substantially reducing prefill and KV-cache costs. The takeaway: don’t just drop redundant tokens. Learn how to fuse them without losing the details.

Yet another paper. This time, we asked a practical question: Do public SKILL.md files actually make agents better? The open skill ecosystem has a quality problem. The hard part isn’t finding more skills. It’s knowing which ones are unique, useful, safe, permissively licensed, and relevant to the task. We crawled ~821K public skills and curated them into 96,401. Then we tested the full retrieval pipeline across 3 benchmarks, 2 harnesses, and 2 open backbones. It improved all three benchmarks, with the largest pooled gain at +7.5pp on SkillsBench. More skills are not automatically better. Curation, retrieval, and the harness determine whether skills become capability.

DeepSeek-V4-Flash is already a first-class model in Raven 🐦‍⬛. Connect your DeepSeek API key and switch in one command: /model deepseek-v4-flash --provider deepseek Your tools, skills, working environment, and EverOS-backed memory stay with the agent.

Raven 🐦‍⬛ v0.1.10 is out. Choose from 21 providers, or search any provider LiteLLM can route to. Recognized deletion commands now ask for one-shot approval in the TUI. Each model call becomes a collapsible episode. Already using Raven? `raven upgrade`

Our new paper. Self-evolving agents have a measurement problem. The hard part isn’t generating changes to an agent harness. It’s knowing which changes genuinely helped rather than got lucky, overfit, or never activated at all. LLMs diagnose failures and propose changes across prompts, knowledge, runtime, tools, and configuration. But deterministic code owns the credit: validity checks, activation checks, paired significance testing, and a sealed test. Across 7 domains, the evolved harnesses gained +9.0 to +15.5pp on the 6 statistically credited sealed tests, retaining 86–147% of the training gain. What transfers is not one universal harness. It’s the diagnose-and-credit loop.

EverOS 1.2.1 is out. LLM-only startup. Add embeddings and reranking later. Three runtime tiers + everos cascade backfill. Also: bounded cascade retries and a security fix for knowledge uploads. On 1.2.0? Upgrade now: `pip install --upgrade everos`

Raven 🐦‍⬛ just crossed 3,000 stars on GitHub. Thank you to everyone who has built, tested, and shared it with us. To celebrate, you can now add this tiny Raven to Codex as your pet.

Raven🐦‍⬛ now supports deep research via @miromind_ai MiroThinker. Some questions should not be answered from memory or the first search result. Raven can now delegate open-ended research, search broadly, cross-check multiple sources, and return a cited report you can inspect later. The image gets the vibe right.

We’re excited to bring @MiniMax_AI to Raven 🐦‍⬛ as a first-class model option. MiniMax powers the model. Raven makes that intelligence compound across runs. Built on EverOS, Raven is a memory-first, self-improving agent harness. It gives models durable user and agent memory, actively manages context during long-running work, follows up proactively through Sentinel, and turns successful workflows into reusable skills that improve with feedback. Paired with MiniMax’s long-context capabilities, agentic reasoning, coding performance, and tool use, Raven can take on longer, more complex workflow, from deep research and coding to automation and persistent digital workers. Try Raven × MiniMax and let us know what you build. Get 12% off MiniMax API plans through our dedicated links:

Update: the EvoAgentBench paper is now on arXiv. This is the paper behind the benchmark we introduced here: measuring whether agents actually improve through experience, ability transfer, and reuse across tasks. Paper:

EverMe × Kimi Code @Kimi_Moonshot @KimiDevs Kimi Code now has shared memory. Connect Kimi Code to EverMe Memory with one prompt, carry context across sessions, agents, machines, and operating systems, and turn what it learns into reusable Skills. Less cold start. More continuity. EverMe: Kimi Code:

EverOS v1.1.1 is live. A small release focused on measurable memory quality: • LoCoMo benchmark runner + docs • Hybrid search optimization (~91% stable) • Agentic search benchmark (~93% stable) • Knowledge cascade upsert race fix • Python 3.12/3.13 CI

EverOS just crossed 10,000 GitHub Stars. ⭐ Big milestone, big achievement. Thanks to everyone who's been part of the journey，the best is still ahead. Stay foolish. Stay tuned.

Meet Raven: a memory-first self-improving agent harness. Powered by EverOS, Raven keeps user memory, agent memory, tools, skills, policies, and execution context together. Successful workflows become reusable agent templates. 🧵

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

EverOS just crossed 9,000 GitHub stars.🎉 The response to our 1.0.0 -> 1.1.0 release means a lot, especially the interest in Knowledge Wiki and Reflections. Thank you to everyone sharing the repo and supporting us. We will keep building agent memory and agentic infra. Thank you. Stay tuned.

Everyone is talking about self-improving agents. The harder question is how to measure whether an agent is actually getting better. That is why we built EvoAgentBench: a benchmark for agent self-evolution. It tests whether agents can learn from past trajectories, extract reusable skills/memory, and improve on held-out tasks across 5 domains: - information retrieval - reasoning and problem decomposition - software engineering - code implementation - knowledge work 917 train tasks. 288 test tasks. In the included Omni-MATH run, skill injection moved 27B from 21% to 65%, and 397B from 25% to 66%. EvoAgentBench has 2K+ all-time downloads on Hugging Face. The benchmark card, task splits, evaluation files, and citation are open there. Self-improving agents need more than vibes. They need measurement.