The 3x YC founder built a staged, Markdown-based knowledge layer for Claude, ChatGPT and Cursor, priced from free to $99 per month.
By RuntimeWire Staff · Published
Primary source: Hacker News
Why it matters #
AI agents are multiplying faster than their memories can interoperate. OzBrain turns that gap into a product, with human-approved writes and portable Markdown as the control layer.
Darius Monsef (@bubs) built OzBrain, giving Claude, ChatGPT, Cursor and other connector-compatible agents a shared place to read and write the knowledge they usually scatter across chats, projects and local files.
The Melbourne-based founder built OzBrain after running into that fragmentation inside his own agent workflow. Monsef says he was orchestrating local Claude Code sessions, remote engineering work, automated routines and planning from his phone while developing an AI voice product for healthcare. Each agent could do useful work, but the results needed somewhere durable to land.
That problem fits Monsef's broader thesis about software. The three-time Y Combinator founder has argued in his writing on agent-operated software that chat and coding agents are becoming the interface through which operators use their tools. OzBrain supplies the continuity required for that interface to survive across models and sessions.
Monsef brings a longer startup record than the typical solo AI-tool builder. He co-founded Creative Market with Aaron Epstein and Chris Williams, and Autodesk acquired the company in 2014. He also co-founded Sightbox before its 2017 acquisition by Johnson & Johnson and ran pediatric urgent-care operator Brave Care from 2019 to 2023. His current biography also lists Calling Round, a conversational AI companion for older Australians, and Monsef Holdings, the Australian entity behind OzBrain.
A wiki maintained by agents
OzBrain stores knowledge as linked Markdown articles rather than an undifferentiated transcript or a collection of retrieved snippets. An agent begins with a compact routing index showing what the brain contains, then opens only the articles relevant to its current task. The examples on OzBrain's homepage include positioning, writing voice, client terms, project status and tool preferences.
The important mechanism is the write path. According to OzBrain's product documentation, an agent stages a proposed change before that change becomes canonical. If the new material conflicts with an existing article, OzBrain s the write and surfaces the contradiction. A user or another authorized agent can review the change before promoting it, while version history records which agent wrote each revision.
That design gives Monsef a credible answer to the mess created when several agents work from the same knowledge. Automatic memory sounds convenient until one model overwrites an agreed price, client term or product decision with an outdated claim. OzBrain makes that conflict visible and keeps the existing version intact until the proposed replacement is approved.
The architecture draws directly from Andrej Karpathy's LLM Wiki blueprint, which described an LLM-maintained collection of interlinked Markdown files. Karpathy argued that agents could take over the bookkeeping that causes human-maintained wikis to decay: updating summaries, repairing links and flagging contradictions as new sources arrive.
OzBrain commercializes that pattern as a hosted connector. The reference on OzBrain's homepage describes inspiration, rather than an endorsement, investment or advisory relationship with Karpathy. RuntimeWire previously covered Karpathy's move to Anthropic's frontier-model research organization, where persistent context remains a separate problem from model capability.
One connector under several agents
Users connect OzBrain through its MCP endpoint. The Claude integration runs through Claude's custom connector flow, while the Cursor integration adds OzBrain as an MCP server. Once authorized, both agents can access the same brain and attribute their writes separately.
OzBrain positions that shared layer below platform-specific features. Claude Projects and Cursor rules files can continue holding instructions tied to a particular workstream or repository. OzBrain is intended for information that must follow a person or team between tools: customer history, operating decisions, preferences, research and the current state of a project.
The distinction matters because AI knowledge portability has become a product question rather than a file-format argument. RuntimeWire reported in July that Formaly founder Arindam Majumder was making a similar case for portable Markdown knowledge. OzBrain extends that premise by letting agents maintain the documents and by putting a review protocol between proposed changes and the accepted record.
Shared brains also move OzBrain beyond personal memory. A collaborator can be granted access so that the collaborator's agents read the same scoped knowledge and submit changes through the same staging system. That could reduce repeated handoffs inside small teams, though it also concentrates sensitive project and client information in a service that every connected agent may be authorized to query.
OzBrain says article bodies are encrypted at rest, accounts are isolated through row-level security, connections can be revoked, and activity can be exported through an audit log. Its privacy policy says Monsef Holdings does not use brain content to train foundation models or sell it for advertising. Content sent to a connected AI service remains subject to that provider's terms.
Monsef is pricing storage, not activity
OzBrain's plan listing shows a free plan with up to 50 articles. Pro costs $20 per month for 300 articles, while Max costs $99 per month for 600. OzBrain says reads, writes, brains and connections are unlimited across those plans, making article capacity the billing constraint rather than agent activity. A custom Company tier is also listed.
That pricing reflects Monsef's core bet: an agent's useful memory will look like an expanding library of maintained work, rather than a metered stream of retrieval calls. It also leaves OzBrain with a demanding trust problem. The material worth sharing across every agent is often the same material users are least willing to expose to an unreliable knowledge system.
Monsef has designed the first version around that concern. Markdown export provides an exit path, staged writes preserve human approval, and provenance shows which agent changed what. OzBrain is currently labeled version 0.1.0, so its ability to keep a large, frequently edited brain coherent will be tested as users move beyond tidy demonstrations and start feeding it contradictory work from several agents at once.
The timing is favorable. Operators are adding agents faster than the platforms are agreeing on a common memory layer, leaving users to copy the same context between products. Monsef is betting that the resulting clerical work can support a standalone product, provided OzBrain becomes the place agents trust before their host platforms make shared memory a native feature.