EverMind launched Raven Agent on July 9th, pitching the product as a self-improving agent harness built on EverOS, its open-source memory operating system, in a PR Newswire release.
The launch is less a conventional chatbot release than a statement of where EverMind wants the agent market to move next: away from one-off sessions and retrieval systems, and toward agents that carry memory across interactions, inspect their own performance, and revise their own operating patterns. EverMind says Raven is available now through its hosted site and on GitHub, while EverOS is available as an open-source project and through a cloud API.
EverMind has not put a named founder or executive at the center of the launch materials. The company describes itself as a global AI company incubated by Shanda Group, which leaves the story unusually institution-led in a category often sold through founder pedigree. The clearest thesis comes from EverMind's own site: "Memory is the defining line between a tool and an agent." That line explains the bet behind Raven more directly than the company's "Digital Life" vocabulary does.
What Raven is supposed to do
Raven is built on EverOS, which EverMind describes as a four-layer architecture made up of an Agent Layer, Memory Layer, Index Layer, and Interface Layer. In the company's telling, Raven turns raw interactions into structured memory units, groups them into contextual scenes, and maintains a profile of the user across identity, preferences, skills, and long-term goals.
The three product claims EverMind is emphasizing are specific. First, Raven is designed for bidirectional memory, meaning it updates a model of the user while also reflecting on its own task performance. Second, EverMind says Raven ships with 100,000 evaluated skills across productivity, professional verticals, and multi-step workflows. Third, EverMind says Raven can rewrite its own skills, runtime logic, and operational strategies, and can work with EverBrain, EverMind's on-device personalized model, to fine-tune model weights.
Those are EverMind's claims, and they deserve to be read as product assertions until developers can test Raven at scale. A system that can revise skills and runtime behavior has obvious appeal for long-running agents, especially in coding, support, research, and personal assistant workflows. It also raises the usual hard questions: who approves a rewritten skill, how regressions are detected, what logs remain inspectable, and where the boundary sits between adaptive behavior and unreviewed automation. EverMind's launch materials do not lay out pricing, enterprise controls, sandboxing policies, or the safety limits around Raven's claimed code-level self-rewriting.
The open-source wedge
Raven's GitHub repository describes it as a self-improving agent harness that provides a runtime, memory layer, tools, and agent templates. The repository says builders can use templates for their own scenario, personality, workflow policy, skills, integrations, or distribution model.
That matters because memory infrastructure is a difficult thing to sell as a standalone abstraction. Developers usually adopt it through a concrete agent, a coding workflow, a support deployment, or a personal assistant that proves the memory layer is worth the operational burden. Raven gives EverMind a showcase application for EverOS, while also turning the harness into a channel for developer-created agents.
The repository outlines a turn-by-turn agent loop with pluggable engines for context, memory, proactivity, tool routing, and evaluation. That modularity lines up with EverMind's launch claim that Raven's memory module, proactivity engine, and tool router are decoupled.
EverMind is also using GitHub traction as part of the launch narrative. The company says EverOS crossed 10,000 GitHub stars within about one month of public release. GitHub currently shows the EverOS repository at 10.7k stars and the Raven repository at 1.3k stars. Star counts are a noisy proxy for adoption, especially in AI infrastructure, but they do indicate that EverMind has found an audience among developers looking for agent memory primitives.
EverOS is the real platform bet
Raven is the product announcement. EverOS is the platform EverMind is trying to make durable.
EverMind's EverOS page describes the system as a portable memory layer for AI agents with persistent memory, multimodal ingestion, and self-evolving skills. The product page says EverOS can ingest PDFs, images, documents, spreadsheets, slides, and URLs in one call, then parse and index them for agent retrieval. It also says memories can be exported as Markdown, and that customers can move from EverOS Cloud to self-hosting under Apache 2.0.
The company's product language is pointed at a real constraint in today's agent stack. Long context windows help, and retrieval-augmented generation helps, but neither fully solves long-horizon continuity. Users still repeat preferences, agents lose procedural lessons between runs, and team workflows often depend on brittle summaries rather than durable operational memory. EverMind's answer is to separate memory from the model and make it inspectable, reusable, and shared across agents.
EverMind backs that pitch with benchmark claims on its site, including 93.05% accuracy on LoCoMo, 83.00% on LongMemEval, and 93.04% on HaluMem. On the EverOS page, the company claims 93%+ retrieval accuracy, under 200ms retrieval latency, and roughly 10x lower cost for its mRAG approach. Those figures are company-published benchmarks; they are useful as a statement of what EverMind is optimizing for, not as independent proof that EverOS will outperform other memory systems inside a customer's production workload.
Raven Builder is the marketplace piece
The next piece is Raven Builder, which EverMind says is forthcoming. The company says developers will be able to define specialized agents for a domain and share them through EverMe, its personal memory hub and digital life management platform. In EverMind's examples, those agents could include a contract law specialist trained by a legal professional or a financial modeling expert cultivated by an investment analyst.
That is the clearest commercial logic in the announcement. If Raven becomes a tool for creating reusable specialized agents, EverMind gets more than a demo for EverOS. It gets a route into a marketplace-like model where agent usage creates more memory data, memory data improves agent performance, and better agents give developers a reason to keep building inside EverMind's stack.
The open questions are the ones that will decide whether that loop becomes real. EverMind has not disclosed the developer terms for Raven Builder, a timeline for its release, pricing for Raven, or the moderation and review model for shared agents. Those details matter more if Raven's defining claim is self-improvement at the skill and code level. A developer marketplace built on adaptive agents needs trust boundaries that are as visible as the templates themselves.
EverMind's launch also formalizes its own ladder for agent maturity. The company describes L1 as role-based functional agents with no persistent memory, L2 as memory-augmented agents with cross-session memory and multi-step planning, L3 as self-improving cognitive agents with reinforcement learning, self-rewriting code, and model fine-tuning, and L4 as autonomous digital life. EverMind says Raven is its bridge to L3.
The taxonomy is company-made, but the strategic point is clear. EverMind is trying to define the agent memory category before larger model providers absorb more of that functionality into their own platforms. Raven gives EverMind a concrete product to test that thesis. If developers use it because the memory loop makes agents more reliable across weeks and months, EverMind has a wedge. If Raven reads as another agent wrapper around impressive but hard-to-verify claims, EverOS will have to win the category on infrastructure merits alone.