Mindstone’s Rebel makes enterprises remember the right model per task, locally

Mindstone’s CTO Greg Detre built Rebel so enterprise AI agents stop “guessing” which model to use, and start remembering it by task. Rebel officially launched this week as a local-first, agentic AI operating system distributed under a “Fair Source” license, and its marquee idea is simple but high-stakes: an organizational memory layer that ensures agents reliably use the enterprise’s preferred AI models for each given task or even subtasks. That routing can switch between local and cloud models in a predictable, visible way, aiming to save costs and maintain data privacy and security as needed.

Rebel is also designed for the real-world constraint that enterprises care about more than demos: inspection and control. Detre describes “Shared memory” as empowering because teams can treat the company as “a super-organism” that gets smarter over time, and Rebel’s memory is meant to be inspectable and portable, not locked in a black box. The system stores state, prompts, task instructions, and a memory hierarchy across local markdown (.md) files, including a primary configuration file called agents.md that acts as the agent’s core instruction layer and runtime boundary.

If you work in operations, security, or platform engineering, Rebel is aiming straight at a common enterprise friction point: orchestration tools can turn into spaghetti. VentureBeat notes that developer-heavy agent frameworks such as LangGraph, CrewAI, and AutoGPT often require teams to wire together databases, cloud infrastructure, and state-management logic. Rebel instead puts the agent memory and instructions in local markdown text files. That shift matters for two reasons. First, markdown can reduce token waste compared with office formats like Word documents and PDFs, because those formats often include formatting and metadata overhead that consumes model context and raises API costs. Second, local text instructions are positioned as a hedge against vendor lock-in. If the instructions, automations, and memory live locally as files, they are less trapped inside a single SaaS provider’s interface or database, which becomes more important as enterprises give AI systems broader access to email, calendars, documents, and internal workflows.

Rebel also adds features that translate from “agent behavior” to repeatable business processes. “Skills” are saved multi-step procedures an agent can reuse. “Operators” let teams adjust how the agent behaves for a given task, such as reviewing a pitch deck from an investor’s perspective or evaluating work through a security lens. “Automations” run scheduled background tasks like scanning messages or files, finding relevant updates, drafting responses, or preparing work before an employee opens the app. Underneath those labels is the same operational bet: enterprises want control over what the system does, how it does it, and what it remembers, without requiring every workflow to start from scratch.

The “best model for each task” capability is where Rebel gets strategically interesting for cost and compliance. Rebel can break a task into parts and route different steps to different models. It can split between local and cloud-based models depending on sensitivity or enterprise policies. The logic is that higher-end models can handle planning or complex reasoning, cheaper models can handle routine work, and local models can handle sensitive steps or approval checks. The source frames this as a way to keep data sovereignty without going all-or-nothing on cloud or local inference. Detre’s explanation ties it directly to enterprise reality: being able to tell the system “Help me with this,” while it knows what is personal, what is sensitive, and what can be shared with the whole company.

Rebel’s memory design also tackles another enterprise AI trap: “dump everything and hope retrieval works.” Instead of pushing large amounts of company information into a database and relying on later search to retrieve the right context, Rebel uses a tiered memory structure. It estimates how likely information will be useful again. High expected value information is written into a local readme.md file tied to a specific project space. Moderate expected value becomes a reference link back to deeper historical records. Lower-priority material is stored in an indexed memory directory, dormant until a relevant task calls it back. This is basically trying to make organizational memory more curated, so agents don’t drown in irrelevant context.

For larger organizations, Mindstone Pro adds an Impact Dashboard aimed at proving where Rebel is saving time and money across business units. Mindstone says the dashboard uses a separate, closed LLM to evaluate telemetry and calculate business impact, and it calibrates conservatively using the lower end of estimated performance gains to avoid inflated productivity claims. That also targets a politically sensitive concern inside enterprises: demonstrating ROI without over-surveilling employees. The source says the dashboard is isolated from individual workspaces, allowing IT and business leaders to evaluate adoption and return on investment without reading employees’ private agent activity.

Finally, Rebel’s licensing is a direct attempt to split the difference between fully closed SaaS and permissive open source. Under the Fair Source license, Rebel’s code is viewable, auditable, modifiable, and deployable. Individuals and organizations with up to 100 concurrent users can run it for free, while organizations exceeding that threshold need a commercial Mindstone Pro license. The license also includes a two-year sunset clause: 24 months after a given version is released, that version automatically converts to the MIT open-source license. For enterprise buyers, the practical implication is risk reduction. If automations, memory files, and agent instructions are stored locally in markdown, a company can move its data and workflows elsewhere if needed. The product debuted on Product Hunt this week, and it prompted technical questions about how the system handles security and related concerns, especially around local approvals and shared memory.