What a knowledge base for AI agents actually is

People use “knowledge base for AI agents” to mean instruction files, automatic memory, vector search, or a reviewable Markdown project brain. Those are different layers. Pick the one that matches the job, then choose a tool.

·5 min read

Someone says they need a knowledge base for AI agents, and they may mean four different things. One team wants a short file the agent reads at the start of every session. Another wants the tool to remember corrections automatically. A third wants to search a large corpus. A fourth wants a current, reviewable answer—why the project chose Postgres, and what was ruled out—so the next session does not start from zero.

Those are not competing brands of the same product. They are different layers. Mixing them up is how an instruction file turns into a diary, or how a personal memory store gets mistaken for a project record.

Instruction files

This layer tells the agent how to work in the repository.

Claude Code’s official memory docs describe CLAUDE.md and AGENTS.md as instructions you write. They are loaded at the start of a session. That makes them useful for conventions, commands, and rules. It also makes them a poor place to keep a running history: every extra paragraph is paid for again tomorrow.

If this file is where your decisions currently live, the question is not “which knowledge base is closest to Notion.” It is how to keep the brief short and put the decisions somewhere else. We have already written about that split in AGENTS.md is not a decision log and AGENTS.md is not a changelog. This page will not repeat those arguments.

Automatic memory

This layer captures what a session learned, without you maintaining a document.

Claude Code’s auto memory is one official example. The same docs say Claude writes notes for itself, scoped to a repository and shared across worktrees, and that the files live under ~/.claude/projects/ on the machine. They are “not shared across machines or cloud environments.” A teammate, or you on another computer, will not see them unless you build another path.

Mem0’s Claude Code integration is another official example, in a different design: capture happens as you work, memories are extracted, and later sessions can recall them. Mem0 describes both shared project memory and personal preferences. That is a retrieval and extraction product, not a Markdown file you review in a pull request.

Automatic memory is the right layer when you want a tool to learn preferences and recent context with little ceremony. It is the wrong layer to treat as the only copy of a decision the whole team must be able to audit.

Search and retrieved context

This layer answers “find the relevant passage” over a large set of docs, tickets, or embeddings. Vector search, enterprise search, and many “chat with your docs” products live here. They are useful when the agent needs to look something up. They do not, by themselves, tell you which retrieved passage is the current decision, or why an older one was reversed.

This page does not benchmark those products, and it does not invent a shortlist. If lookup is the job, choose a search tool on its own terms.

A reviewable project brain

This layer is the current answer, plus the trail that produced it, in a form a human or another agent can read.

If a fresh session cannot see it, it is not project context yet. A decision the next session never saw is often a meeting or a chat that never became a record. A memory only one agent can see is not project memory yet. The shape that addresses those failures is a project brain: a current conclusion that can be rewritten when the truth changes, and an append-only history of how you got there.

BRAIN.md is an open, local convention for that shape. A BRAIN.md file is the protocol entry point. The brain/ folder holds root documents and pages; pages keep a compiled_truth and a timeline. It is git-native Markdown, not a hidden index. It is also not AGENTS.md. The instruction file remains the brief; the brain holds the decisions.

Letta’s Context Repositories are a related official design in this neighborhood: memory stored as files, versioned with git, so agents and humans can inspect and update it. The existence of more than one approach is the point. A reviewable, file-based memory is a category, not a single vendor.

MindMux’s Project Brain is the product form of this layer: readable project knowledge your coding agents can share across sessions and tools. It does not claim to replace instruction files, automatic memory, or vector search, and it does not claim to ingest GitHub, Slack, or email automatically.

Pick the layer, then the tool

If you need the agent to follow rules, write a short instruction file. If you need it to remember preferences or recent corrections, use a memory system that is honest about where that data lives and who can see it. If you need to search a corpus, use search. If you need the next session to read the current decision and the reason behind it, you want a project brain.

When that last job also needs the cloud product, create a MindMux account. After you have an account, the recommended way to connect Codex or Claude Code is the MindMux CLI over MCP. Installing the CLI is the connection step, not the account. If you only want the local convention, start with BRAIN.md and keep the files in git.

If you arrived here from a more generic search, the same split applies. A Notion alternative should be chosen by the job, not by the closest clone. Notion AI is a workspace assistant; it is not automatically a project brain. An Obsidian alternative is sometimes a new vault, and only sometimes a shared source of truth for agents.

Share to