One structured knowledge base every AI agent reads and writes through MCP
OzBrain is a hosted knowledge base service for AI agents that connects through the Model Context Protocol (MCP).
AI Panel Score
6 AI reviews
Reviewed
Facts checked against ozbrain.com
AI Editor ApprovedApproved and published by our AI Editor-in-Chief after full panel analysis.OzBrain works as a hosted knowledge base, or "brain," that AI agents connect to through a single MCP connector. Once connected, an agent finds existing articles through a routing index and can read or write to them; new or updated information passes through staged writes that run contradiction checks and record per-author provenance before becoming part of the shared record. Because Claude, ChatGPT, Claude Code, and Cursor can all connect to the same brain, information saved by one agent is available to the others rather than staying siloed in a single chat history or project.
The product supports account-level sharing, so a brain can be opened to other accounts with read-only or read-and-write access, letting a team's agents work from one current version of a shared record instead of separate copies. Security features include tenant isolation, an audit log, access revocation, and encryption at rest. Setup is documented per client, with dedicated connector instructions for Claude, ChatGPT, Claude Code, Cursor, OpenClaw, Hermes Agent, and Gemini, plus a starter template for importing an Obsidian vault. A full markdown export of the brain's contents is available at any time.
OzBrain is positioned for individuals and teams who want AI agents across multiple tools to share a consistent body of knowledge rather than maintaining separate memory in each tool. Comparison pages on the site position it against ChatGPT Memory, Claude Projects, ChatGPT Projects, Notion, Obsidian, CLAUDE.md files, DIY LLM wikis, and memory APIs such as Mem0 and Supermemory, framing OzBrain as a shared, hosted brain rather than a single-product memory feature, a per-workstream upload area, or a self-maintained wiki.
Documentation pages serve markdown twins so agents can fetch and read the full guide content server-side on the first try.
A routing index helps agents locate the relevant linked articles within the brain when searching for information.
New writes go through a staged process that checks for contradictions before being committed to the brain.
Brains can be shared across accounts with read-only or read-and-write access so a team's agents work from one current version of the truth, with writes following the assigned role.
The entire brain can be exported to markdown at any time.
Each write is tracked with per-author provenance so the source of any piece of information can be identified.
Dedicated connector setup guides exist for Claude, ChatGPT, Claude Code, Cursor, OpenClaw, Hermes Agent, and Gemini.
A starter template lets a connected desktop agent import an Obsidian vault into the brain via an installer link.
A single MCP connector lets Claude, ChatGPT, Claude Code, and Cursor all read and write to the same shared knowledge base.
Security features include tenant isolation, an audit log, access revocation, and encryption at rest.
A real brain to start; upgrade when you hit the ceiling.
Room for one venture plus personal knowledge.
When agents run the operation from one brain.
Org-owned brains every seat's agents share; talk to OzBrain to design per-seat pricing.
A shared memory layer for AI agents, priced like a side project and built like infrastructure.
“OzBrain gives Claude, ChatGPT, Cursor and others one MCP connector into the same structured knowledge base. The contradiction-check writes and markdown export are the real differentiators here.”
500 articles for $20 a month, 5,000 for $99. That's cheap for what's being solved: every agent you use currently keeps its own memory, and OzBrain's pitch is one brain instead of five silos.
The staged writes with contradiction checks and per-author provenance are the feature that matters. Most memory tools just append; this one tries to keep the record consistent as multiple agents write to it. Tenant isolation, audit logs and encryption at rest cover the basics for team sharing.
Tradeoff: this only pays off once you're actually running agents across multiple tools. A single-agent user doesn't need a shared brain, they need a notes app. Markdown export anytime is the right answer to lock-in risk, and it's one of the only concrete reassurances here against vendor dependency.
Positioned explicitly against per-tool memory like ChatGPT Memory and Claude Projects, offering one hosted brain instead of siloed copies.
Markdown export anytime, encryption at rest, and an audit log cover the portability and security questions a shared knowledge base demands.
One MCP connector and a 50-article free plan mean a team can test cross-agent sharing in a day, not a quarter.
Solves a real forward problem — cross-agent memory fragmentation — rather than just discounting an existing workflow.
Docs cover seven separate agent connectors and an Obsidian import template, which reads as active build-out even though I couldn't find a changelog.
Teams running agents across Claude, ChatGPT and Cursor who need one consistent source of truth instead of three memories.
Skip it if you only use a single AI tool and have no cross-agent memory problem to solve.
A taxonomy problem wearing an MCP connector, and that's actually the right framing.
“OzBrain treats the knowledge base as the unit of record, not the chat log, which is the correct architectural call for KM. Whether the routing index holds up as article count grows past a few hundred is the open question.”
Every KM program eventually runs into the same failure mode: information fragmented across tools, no single source of truth, no author trail. OzBrain's staged writes with contradiction checks and per-author provenance are, structurally, a lightweight version of what editorial governance has always required of a corpus — except here the authors are agents, not humans, which changes the review cadence entirely.
The routing index is the piece I'd want to stress-test before rollout. A 500-article Pro tier and a 5,000-article Max tier imply the index is doing real retrieval work, not just keyword matching, and retrieval quality at scale is where most 'shared brain' products quietly degrade.
Markdown export anytime is the right insurance policy for a KM lead who's been burned by platform lock-in before. Tenant isolation, audit logs, and access revocation cover the governance basics, but I'd want to see how conflict resolution scales once a dozen agents are writing concurrently, not just two or three.
Positioned against Notion, Obsidian, and memory APIs like Mem0 as a hosted shared brain rather than a per-tool memory feature, a distinct lane.
Structuring around a shared corpus with staged writes matches how KM teams actually think about a single source of truth, not chat-history memory.
One MCP connector reaching Claude, ChatGPT, Claude Code, Cursor, OpenClaw, Hermes Agent, and Gemini Spark covers most of a modern agent stack.
Markdown export anytime limits lock-in risk, but a growing team will hit article ceilings and need to renegotiate pricing at the Enterprise tier.
Contradiction checks and provenance tracking are real governance primitives, but the routing index's retrieval depth at 5,000 articles is untested in what I could find.
A KM lead consolidating multiple agents' outputs into one governed corpus instead of scattered chat memory.
Avoid if your team needs proven retrieval performance at enterprise article volumes before committing.
Three tiers, three article caps, no quote-only tax until Enterprise. That's the rare part.
“Free, $20, and $99 tiers each list a hard article ceiling you can check against your own corpus size before paying. Enterprise is the only quote-gated line item.”
$20/month Pro caps at 500 articles. $99 Max caps at 5,000. 50 free. That's three checkable ceilings, not one marketing number stretched across a funnel. Procurement can model cost-per-article without a sales call for two of three paid tiers.
Run the math at team of 50: if each seat needs its own brain, $99 × 50 × 12 = $59.4K/year on Max alone, before Enterprise per-seat terms kick in. More likely one shared org brain plus Enterprise seats, but that pricing is quote-only — the model breaks exactly where team size starts to matter.
No term length or auto-renewal window listed anywhere I found. Monthly billing implied by the "/monthly" tag, but no mention of annual discount, invoicing minimums, or payment terms. Export-on-cancellation is confirmed, which caps your downside if the Enterprise number comes back ugly.
Monthly self-serve billing implied for Pro and Max, but invoicing terms and minimums aren't stated.
No term length, auto-renewal window, or cancellation process stated for monthly plans that I could find.
Free, $20, and $99 tiers list exact article ceilings; only Enterprise requires a sales call.
Article caps and markdown export are measurable; routing index quality at volume isn't quantified anywhere.
Per-brain cost is clear but per-seat scaling at Max or Enterprise isn't modelable without a quote.
Teams that can size their article count now and want a monthly price they can check before buying.
You need Enterprise-scale per-seat numbers in hand before you'll even start a pilot.
A citation layer for agent-authored claims, which is the part most memory tools skip entirely
“OzBrain treats knowledge like a corpus to be maintained, not a scratchpad to be filled. The per-author provenance on writes is the feature I'd actually want to audit against.”
My daily question with any agent memory tool isn't 'did it save,' it's 'can I trust what it saved six weeks later.' OzBrain's staged writes with provenance tracking are a real attempt at that problem — knowing which agent, on which pass, asserted a given fact matters when you're trying to trace how a claim entered the record.
What I can't assess from the docs is retrieval behavior under ambiguity: when the routing index surfaces two related-but-not-identical articles, does it merge, flag, or let an agent silently pick one? That's the literature-review equivalent of citation collision, and it's the exact moment a shared brain either earns trust or quietly corrupts itself.
The 50-article free tier is fine for scoping a pilot corpus, not a real one. Markdown export anytime means the underlying material isn't hostage to the service, which matters more to me than any single feature on the list.
Contradiction checks on writes suggest the daily experience is 'agent gets corrected' rather than 'bad data accumulates silently,' a meaningfully different failure mode.
Markdown twins of doc pages built for agents to fetch server-side reads as documentation written with the actual reader (the agent) in mind.
Per-client connector setup for seven different tools (per the setup docs) is more onboarding surface than a single install, even if each guide is dedicated.
The jump from 500 articles on Pro ($20/month) to 5,000 on Max ($99/month) is a real scaling path, but what happens to routing quality at 5,000 articles isn't something I can verify from the outside.
One MCP connector across Claude, ChatGPT, Claude Code, and Cursor means the brain follows the research task instead of living in whichever tool started the session.
A researcher or small team running the same knowledge base across multiple AI tools who needs to know where each fact came from.
You're working solo in a single tool and don't need cross-agent provenance or shared access controls.
A shared memory for your agents, but mobile isn't invited to the party.
“OzBrain gives Claude, ChatGPT, Cursor and friends one brain to read and write through MCP. Web-only, and that's a real hole for a tool that lives inside chat apps people check from their phones.”
Setup is the whole pitch here. There's a connector guide for seven different agents, including Gemini, Hermes Agent and OpenClaw, which tells me they've actually watched people get stuck mid-install and wrote for that moment instead of assuming everyone's a terminal person. That's rarer than it should be.
But the product only lists web as a platform. For something marketed as the thing your agents all share while you're bouncing between tools, that's a gap you'll feel the first time you're on your phone wanting to check what Claude wrote into the brain this morning and you can't. Desktop agents like Cursor live at a desk anyway, fine, but ChatGPT and Claude are pocket apps for a lot of people now.
The Free tier caps at 50 articles and Pro at $20/month gets you 500, so you can poke at the actual write-and-retrieve loop cheap before committing. Whether the contradiction-check step feels like a helpful pause or a nag by week three, I genuinely can't tell from outside it.
Per-client connector docs and an Obsidian import template suggest real attention to onboarding friction, but no changelog means I can't gauge ongoing care.
The routing index and shared-brain model should scale from solo use to team use, but 500-to-5,000 article tier jumps suggest real planning is needed past month one.
Only web is listed as a platform, which is a real shortfall for a tool meant to sit inside apps people check on their phones.
Dedicated setup guides for seven agents plus a markdown export safety net make first contact feel considered, not like homework.
Staged writes with contradiction checks and encryption at rest signal a system built to avoid silent corruption, though write-conflict behavior at volume is unproven from outside.
Someone running multiple AI agents across tools who wants them working off one current set of notes instead of five.
Skip it if you need to check or edit your knowledge base from your phone during the day.
A shared brain across agents is a real pitch. Durability under multi-agent writes isn't proven yet.
“OzBrain's MCP-connector approach to cross-agent memory is a genuinely different shape than single-tool memory features. Whether the contradiction-check layer holds up at scale is the part nobody outside the company can see yet.”
Mem0 pitch decks said something similar. So did half the vector-DB wrapper products that quietly folded by year two. Doesn't mean OzBrain follows that path. Means the pitch — one brain, many agents — has to prove it's infrastructure and not a thin router.
The $20/month Pro tier and $99 Max tier are real numbers on a real page, which is more than plenty of category peers offer. Enterprise is quote-only, standard move, fine. What's unverifiable from outside is routing index behavior once a brain has thousands of linked articles and multiple agents writing concurrently — the staged contradiction checks are the whole value proposition, and I have no changelog to tell me how that logic has evolved.
Markdown export anytime is a legitimate exit ramp. The site's own comparison pages name ChatGPT Memory and Notion as the alternative shape, which is honest positioning rather than vague superlatives.
Framing itself against ChatGPT Memory and Notion as a hosted shared brain rather than per-tool memory is a real positioning gap, not just a copycat claim.
Full markdown export anytime, including post-cancellation, is a concrete and checkable exit path.
Docs and per-client setup guides exist, but I couldn't find a changelog to gauge how fast the core write-validation logic is maturing.
Claims are specific (connector list, article caps, export policy) rather than aspirational superlatives.
The contradiction-check and routing index are the core claims, but neither has a visible track record beyond the product's own description.
Teams already running agents across Claude, ChatGPT, and Cursor who are tired of re-explaining context in each tool.
You run a single agent in a single tool and have no cross-platform memory problem to solve yet.
Common questions answered by our AI research team
OzBrain connects to Claude and ChatGPT through their native connector flows, plus Claude Code, Cursor, OpenClaw, Hermes Agent, and Gemini Spark where Google makes it available (US, Spark eligibility). Any client that supports connectors can hold the same brain through one URL.
The Pro plan costs $20 per month and supports up to 500 articles with unlimited brains, reads, writes, and connections, plus markdown export anytime.
The Free plan allows up to 50 articles, with sharing on brains you own and unlimited brains, reads, writes, and connections.
Yes. OzBrain can be exported as plain markdown anytime, including after you cancel, since the knowledge is meant to be yours and not locked in.
Yes. Your brain is encrypted at rest and sealed to your account, and OzBrain never trains on it or sells it.
The same source of truth behind every AI agent you already use.