One shared, persistent memory for all of your AI agents. What one agent saves — your work, your own knowledge, the shape of a recurring task like particular paperwork or particular invoices — is there for every agent you connect, across sessions and across machines, so your AI does not have to make the same mistakes over again.
Your agents connect to Vanilla Brain over MCP (the Model Context Protocol).
Setup installs a short memory protocol that asks each agent to call
brain_wake at the start of a session, pull relevant context with
brain_recall, and save durable facts with
brain_remember. Everything is one account — connect Claude
Code on one machine and Codex on another, and they share the same memory.
A recall finds more when it is asked more than one way.
brain_recall takes up to two other wordings of the same question
in alternatives and searches all of them in one call; each is
ranked on its own and the lists are merged. Word each one the way a saved
memory would put it (the name of the thing, the tool or the person) rather
than in the words of the question.
We process your content only to deliver the features you authorize. We do not use it to train AI models, sell it, use it for advertising, or routinely read it. The hosted preview processes readable content; limited authorized human access remains possible for essential support, security, and legal obligations. Your chosen AI agents receive the context you authorize under their providers’ own terms. Read the complete Terms & Privacy Policy for the access limits, retention, and deletion process.
Guided setup currently supports Claude Code and Codex.
brain_status, brain_wake and a focused
brain_recall. Setup text alone is not proof of a working
connection, and a new connector in Settings > Connections does not by itself show
that it came from this file.Conversation capture is off until you approve it. When you do, it runs as Claude Code or Codex hooks on macOS and Linux and records available visible prompts and final responses. Missing or refused text is represented by content-free gap records. Hooks never read or upload hidden reasoning, system instructions, tool traces or incidental documents.
Optional sources and current continuity limits. In the dashboard, under Settings > Memory, you can enable a checklist of the requests you type. The checklist is off by default. Automatic Claude Code compaction capture and automatic continuity injection are currently unavailable: native hooks can copy parent routing without dependable child identity. The retained compaction source authorization does not activate intake. Old queued summaries stay local for review. A transcript-recovery setting is reserved: no capture hook in this release reads transcripts. Codex compactions are encrypted by the provider and are never collected. These settings are yours alone: they apply from the moment you change them, never to anything earlier, and never to another account.
Documents. Your AI selects documents worth keeping (an authoritative source, a lasting decision, an operating instruction or a final deliverable) and records the exact revision, its provenance and why it matters. By default you approve each one. You can instead let the connectors you choose keep the documents they select without per-document review, within your exclusions and a daily limit. Either way it is selective: files are never collected merely because an agent read or edited them, and you can inspect, pause, revoke and delete what was kept.
Each setup file adds at most one connector entry in Settings > Connections; its capture-only credential is managed with it, and revoking the connector revokes both. Every connector on your account reads and writes the same memory.
Other MCP clients can connect manually with a key created in Settings > Connections. A manual key has full access, including deleting memories, so prefer a setup file where you can.
Vanilla Brain never pushes to an agent. On Claude Code and Codex you can install hooks that let the agent’s own runtime bring memory in at the right moments: the wake briefing when a session starts, a focused recall on its first real prompt, and new messages from your other agents (the agent mesh) while it works, and a bounded reminder if a request is still unanswered: on Claude Code before it stops, on Codex on the next prompt. The hooks reuse your existing connection, fail open in about two seconds when the service is unreachable, and back up your settings before merging into them. Ask your agent to install them, or run:
dir="$(mktemp -d)"
curl -fsSL https://vanillabrain.ai/hooks/vb_hook.py -o "$dir/vb_hook.py"
python3 "$dir/vb_hook.py" installpython3 ~/.config/vanilla-brain/vb_hook.py uninstall removes
only these hooks. Codex asks you to review and trust new hooks in
/hooks before they run.
Night Shift, the optional memory tidy-up, is set up in chat. Ask your connected agent, for example: “Set up Night Shift to use 10% of my usage and run it every night at 2.” The share of your usage is the only setting. The agent can start or stop a run, schedule it as a Claude Code scheduled task or a Codex automation, or install the hourly runner.