The open-source projects worth knowing for a company brain
Anyone who decides to build a company brain rather than buy one finds the same thing: there is no single repository to clone. There are four categories of project, each solving a different part, and a build means picking one from two or three of them and writing the glue yourself.
What follows is sorted by what each project does rather than by popularity. Star counts and licences were read from the GitHub API on 30 August 2026 and will have moved since; they are here as a rough signal of whether a project has a community behind it, not as a ranking.
The opinionated whole
GBrain is the closest thing in public to one person's finished answer. Garry Tan, president of Y Combinator, built it to run his own agents, and its README describes the corpus it holds for him: roughly 155,000 pages, 24,500 people, 5,300 companies, with cron jobs that ingest and consolidate overnight. It has since grown a team mode where each person queries their own slice, scoped by login.
Two things make it worth reading even if you never run it. It answers rather than retrieves, synthesizing cited prose instead of returning ten chunks, and it states explicitly what the brain does not yet know. And it builds a typed knowledge graph as a side effect of every page write rather than by calling a model, which is what lets it answer "who works at Acme" at all. Its benchmarks live in a separate evals repo, which is more accountability than this category usually offers.
| Project | What it is | Licence | Stars |
|---|---|---|---|
| GBrain | An opinionated personal and team brain with synthesis, a self-wiring graph, and an overnight consolidation cycle | MIT | 29.3k |
Memory layers
The problem these solve is narrow and real: an assistant forgets everything between sessions. They sit under an agent and persist what it learned.
They are not company brains on their own. Memory is per-agent or per-user by default, and turning that into something a whole company shares is the work you would be signing up for.
| Project | What it is | Licence | Stars |
|---|---|---|---|
| Mem0 | A universal memory layer for agents, the most widely adopted of these | Apache-2.0 | 64.3k |
| Letta | Stateful agents with long-term memory, out of the MemGPT research | Apache-2.0 | 24.5k |
| Cognee | Memory for agents, built around a graph rather than a document store | Apache-2.0 | 30.3k |
| Basic Memory | Conversation memory kept as plain Markdown files you own | AGPL-3.0 | 3.8k |
Basic Memory is the one to look at if the plain-files, version-history model is what appeals. It is much smaller than the others and that is the point.
Search and chat over your own documents
The category most people mean by "internal AI". Connect the sources, index them, ask questions in a chat window.
This is retrieval, with the limits retrieval has: you get ranked passages rather than a reconciled position, and undated passages at that. For finding a document it is exactly right.
| Project | What it is | Licence | Stars |
|---|---|---|---|
| Onyx | Team chat and search across connected company tools, formerly Danswer | Source-available | 31.8k |
| Khoj | Self-hostable assistant over your documents and the web, with scheduled automations | AGPL-3.0 | 36.8k |
| AnythingLLM | Local-first workspace over your own documents | MIT | 65.4k |
| RAGFlow | A retrieval engine built around deep document understanding | Apache-2.0 | 89.6k |
Graphs and frameworks
If you are building rather than deploying, these are the parts. The graph projects exist because ranked passages were not enough on their own, which is the same conclusion a company brain starts from.
| Project | What it is | Licence | Stars |
|---|---|---|---|
| GraphRAG | Microsoft's graph-based approach to retrieval over a private corpus | MIT | 35.7k |
| Graphiti | Real-time knowledge graphs for agents, with time built in | Apache-2.0 | 30.4k |
| LlamaIndex | The document and ingestion layer most of these are built on | MIT | 51.9k |
| Haystack | Orchestration for production retrieval pipelines | Apache-2.0 | 26.4k |
| Dify | A workspace for building agentic workflows and RAG pipelines | Source-available | 153.9k |
| MCP servers | Reference servers for the protocol assistants use to reach outside systems | Mixed | 90.0k |
Graphiti deserves a specific note: it holds when a fact was true, not just that it is related to another fact. That is the property most retrieval stacks are missing, and the one that decides whether an assistant can tell this quarter's pricing from last year's.
What you still have to build
Clone any of the above and you have the mechanism. What you do not have is the part that makes a company brain trustworthy, and it is the same short list every time.
Keeping the pages true. Every project here will index what you point it at. Almost none of them will notice that a process changed and revise the page that described it. That loop is the difference between a brain and a snapshot, and it is where the ongoing work lives.
Deciding who sees what. A brain assembled from systems with different access rules, that does not carry those rules, is a data incident with a good interface. GBrain's per-login scoping exists for exactly this reason, and it is worth noting how few of the others have an answer at all.
Somewhere to run it. Most of these are a service, a database and a model provider, all of which are yours to operate and pay for.
Build or buy
None of that is an argument against building. It is what the build actually consists of, past the clone. If your team has the appetite, start from the category that matches the part you care most about and be honest that the glue is the project. If it does not, that is the case for buying one, and what a company brain is is the shape to compare against.
Common questions
- Is there one project that is a whole company brain?
- Not really. GBrain is the closest thing in public to a finished opinionated answer, and the rest of the field is parts: memory layers, search over your documents, knowledge graphs, and frameworks to assemble your own. A build means picking from two or three categories and writing the glue.
- What does building one actually cost?
- Past the clone, it is a service, a database and a model provider to operate and pay for, plus the two things almost none of the projects give you: the loop that notices a process changed and revises the page describing it, and access rules that follow your teams rather than ignoring them.
- Are the star counts and licences current?
- They were read from the GitHub API on 30 August 2026 and are stamped with that date in the article. They will have moved since, and they are here as a rough signal of whether a project has a community behind it rather than as a ranking.