Open source keeps a company brain in files you own
A company brain is the memory your agents read and write. Four kinds of product sell one; open source keeps the files in a repo you own.
A company brain is the memory your agents read at the start of a task and write back to when they learn something. Slite puts the working definition this way: "A company brain is a persistent context layer. It gathers knowledge from wherever a company creates it, and it makes that context usable by both people and AI agents" (Slite). Each vendor defines it as the product it sells, which is why the definitions have not settled. Kortix is open source, and it keeps a company brain as Markdown files in folders you own, with the repo copy landing through a change request a person merges.
A brain is memory, and memory sits somewhere
Kortix is the open-source CompanyOS: your agents, their skills, your company memory, and every connector in one git repo you own, with the agents working on real cloud computers.
People cover for missing context without noticing: someone asks a colleague where a decision came from. An agent can neither ask nor remember. It knows only what it was handed, and forgets when the session ends. Slite's guide makes the point well: context sprawl was a nuisance until agents started doing real work, and then it became a dependency.
The decision that lasts is who holds the files and who approves a write.
The term is about a year old
Four threads met inside twelve months, as Slite's guide tracked the phrase. Foundation Capital published its piece on context graphs on 2025-12-22, arguing that most systems keep a fragment of a project and none the reason it happened. Andrej Karpathy proposed an LLM wiki in April 2026. Garry Tan open-sourced GBrain for his own agents. Y Combinator's Summer 2026 Request for Startups carried an entry called Company Brain by Tom Blomfield, which cited GBrain and asked for the same thing for every business. A vendor picks whichever thread suits the product, which is why the definitions disagree.
What each kind of product is for
Four kinds of product answer "company brain", and they are not four versions of one.
| Tool | In its own words | Kind of product |
|---|---|---|
| Kortix | "Memory is Markdown files in Drive", plus memory/ in the repo | A brain in a repo you own |
| GBrain | "Give the agent you already use a memory you control" | A brain in a repo you own |
| Glean | "Enterprise AI that understands your company" | Knowledge base and enterprise search |
| Guru | "The Governed Knowledge Layer for Enterprise AI" | Knowledge base and enterprise search |
| Dust | "Multiplayer AI for human-agent collaboration" | Knowledge base and enterprise search |
| Mem0 | "Drop-in memory infrastructure for AI agents and apps" | Memory infrastructure a developer embeds |
| Sentra | "The managed organization memory" | Managed organizational memory |
The knowledge base group is the crowded one. Guru says it "structures, governs, and continuously improves your company's knowledge"; Glean calls its answers "secure, explainable, and permission-aware"; Notion AI answers inside the workspace. Tettra builds an internal knowledge base from Google Docs and Notion; Confluence is the wiki many teams already run; Elium sells knowledge "kept current by its owners"; Bloomfire turns enterprise information into "a governed reasoning layer for every AI source and agent". Slite and company-brain.ai sell the same layer as context, Slite through a guide that defines the term and company-brain.ai through an MCP connector. Dust connects 70+ sources so agents "answer with the full picture and take action".
The developer tools are a different animal. Mem0 ships "drop-in memory infrastructure"; Cognee connects to Claude Code, Cursor or any MCP client; Letta researches agents that "remember everything"; Graphiti is "a framework for building and querying temporal context graphs for AI agents"; LangMem "helps agents learn and adapt"; Pinecone sells retrieval that "doesn't get more expensive as it scales". None holds the company's record, only what one product's agents have seen. Each group answers a different question:
- Who may read it, and is it current. The knowledge base question, and the one most companies have.
- Who approves a change, and can it be acted on. Sentra reads ERP, CRM, HR and email, raises what looks wrong, and acts once someone says yes. Buy it when nobody wants to run a pipeline.
- Where does what the agent learned go. The memory library question. Buy it when memory is a feature of the product you are building, not an asset the company owns.
- Who holds the record, and who signs off on a write. The repo question, the one Kortix and GBrain answer, and the one Kortix is built around. Buy it when the memory has to outlive the vendor.
The repo route is a standing job: somebody reads the diffs. A buyer who picks from a demo picks the group whose question they had.
What an agent needs from memory
Three requirements show up in every setup that works, and chat history fails all three.
- Read at the start of a task. Memory pasted into a prompt is a guess about what this task needs. Kortix does not paste memory into every prompt: a built-in skill tells each agent to read memory at the start of a task (docs).
- Survives a context reset. The starter's worker agent puts it plainly: "Assume interruption: your context can reset, and only what's written to
memory/survives" (starter worker). - Scoped to who may read it. A brain every member can read is a leak with good search.
Where the files live in Kortix
Every Memory folder is read with its folder's rights. Company memory lives in the Company folder, where every project member reads it; your own Memory folder opens only to you, with each agent keeping notes in a subfolder; a channel's Memory folder opens to that channel's members, when Channels is on. Drive is experimental today.
The repo copy is a memory/ folder with MEMORY.md as its index. The built-in memory tool edits it, and every edit lands through a change request (docs). Every new project also ships the harness-reflector agent: it reviews recent sessions, then improves skills, instructions, tools and memory, bundling the work into one harness: … change request it never merges. Its daily schedule ships switched off, and turning it on is one line in the project config, merged like any other change. Merge is default-deny for agents.
Slite's argument for a managed brain
Slite's guide says to choose a personal or open-source tool "when the memory is yours or your team is small and close-knit", and a managed Company Brain like Slite "when access levels differ, sources disagree, sensitive data is involved, or agents can take actions". It is the strongest case against keeping the files.
Access levels differ, and folder rights cover that without a vendor in the middle. Sensitive data stays out: "Never put passwords, API keys or personal data in memory. Store credentials as project secrets instead." Agents taking actions is the clause that matters, and each connector action carries an Allow, Ask or Block rule, with Ask holding the call until a person decides.
Where Slite wins is real. Sources that disagree do not reconcile themselves, and a managed product is less work in the first month. The files then sit in someone else's product, and leaving means trusting that vendor's export. If nobody on your team wants to read those diffs, buy the managed brain.
Four questions for any company-brain vendor
- Where do the files live, and can you export them in a format you can read without the product?
- Who reads each scope, and does a new folder inherit the folder's rights?
- Who approves a write, and can the agent that proposed it also approve it?
- What survives a change of vendor: the facts, the index, or only the chat history?
Kortix answers with a repo you can clone, three Memory folders and the change request on memory/. Run it on Kortix Cloud, in your VPC, or on your own on-prem network. Self-host is free.
Next reads: Open source is what makes the stack yours and The test of sovereignty.
Start with one agent and one memory folder, then read the diff it opens. Free to start, free to self-host. Try Kortix