Frequently asked questions

Hebby is a long-term memory and second brain for AI agents like Claude Code and Codex. It remembers what your model can't know on its own: your decisions, projects, systems and pitfalls. And it brings those memories into the session exactly when they're needed. Your agent never starts from zero again.

Hebby is modeled on human memory and built by psychologists. Instead of just finding similar text, Hebby thinks in connections: an in-house Graph Memory links your knowledge, Hebbian learning strengthens what is used together, and Hebby anticipates which memories will matter next.

Quite the opposite. Hebby complements the model instead of competing with it. Better models know more about the world, but never more about you. And a stronger model gets even more out of the same memories. Hebby grows with every model update.

When neurons are active together, their synaptic connection grows stronger. The brain learns connections by this principle, described by the psychologist Donald Hebb. Hebby applies it to AI: what is used together in your work, Hebby links more closely. So your memory gets smarter with every session. More on the page Hebbian learning.

Yes, one that thinks along. Classic second-brain apps like Obsidian or Notion need to be maintained by you. Hebby fills and organizes itself on its own while you work with your AI agents. More in the guide Second Brain with AI.

No. A context window is working memory: once the session ends, everything is gone, and the fuller it gets, the worse the model performs. Hebby is long-term memory and brings only what matters. More in the guide Context window explained.

Hebby is built by psychologists in Germany, based in Hamburg. How people remember, forget and form connections is the blueprint for Hebby. More on About Hebby.

The brain's memory center is called the hippocampus, so a hippo was the obvious choice. The name combines Donald Hebb, the father of Hebbian learning, with “happy memories”: Hebb + happy memories = hebbier memories. And who could forget a hippo in a doctoral cap?

Hebby is ready-made for Claude Code and Codex and switches itself on automatically in every session there. Other agents and tools can use Hebby via the Model Context Protocol (MCP).

Yes. What Claude Code learns on your laptop, Codex on your server knows too. One memory for all agents, machines and team members.

No. Hebby notes what matters while you work and organizes its memory on its own, like a brain during sleep. Outdated things recede, important things stay present. In the portal you can see what Hebby knows at any time.

Yes. Measured in live operation, Hebby costs 3 to 10% additional tokens depending on usage, about 5% on average. Long automated loops are at the low end, lots of dialogue and brainstorming at the high end. In return, research, follow-up questions and repeated mistakes disappear, and context files like CLAUDE.md or AGENTS.md stay almost empty. Save just two work steps per request and Hebby is already in the black. See the measurement method and calculation.

No. Hebby comes with a built-in, lean reranker, but doesn't depend on it. Hebby learns on its own which memories belong to which work. The longer it runs, the more accurate it gets by itself, and the less often the reranker is needed. More in the guide Reranker explained.

By the strictest criterion: used. A memory only counts when it is in context at the right moment and actually changes a step of your agent. The goal is at least one used memory per request, because that makes every run better. Hebby achieves that on practically every request. The four quality levels.

Hebby measures on its own. Built-in statistics show you at any time what your memory costs in tokens and what share that is of your total AI costs. Full cost control, no guesswork.

Hardly. We recommend leaving context files like CLAUDE.md, AGENTS.md, GEMINI.md or MEMORY.md almost empty. Only what must apply in every single session without exception belongs there, usually a sentence or two. Hebby handles everything else and brings it exactly when it's needed. That saves tokens on every prompt and the constant tidying of those files is gone.

We recommend it. Built-in memory features like memory files or auto-memory can keep running alongside Hebby, but they cost extra tokens and often say the same thing twice. Hebby is built to take over this job completely. So turning them off saves even more.

No, even if AI assistants like to claim so. A second brain doesn't only pay off for huge projects: it delivers better results, scales far better than built-in memories or homemade memory files, and lowers token costs per prompt from day one.

For everyone who works with AI agents every day: developers, teams, agencies and freelancers who don't want to explain everything from scratch each day.

Hebby is developed in Germany, and the hosted version runs in a data centre in Frankfurt, Germany. Your memory belongs to you: you can see what is stored at any time and delete entries. Passwords and credentials generally don't belong in Hebby.

Only the agents and machines you approve. You have to explicitly confirm new connections, and Hebby controls who may read what.

Hebby is handed out in Early Access to selected users. Request access via the Early Access page or write to office.weihrauch@gmail.com.

Hosted from €5 per month, for teams €9 per person. Or a one-time €199 to self-host on your own server. Anyone who joins in Early Access keeps their price permanently. All pricing.

No. Hebby uses your own AI subscription or API key. We add no markup to your tokens and don't earn from your usage. You only pay for the memory.

Hosted, Hebby runs on our servers in Frankfurt, Germany. You get your access and are ready to go in minutes, and we take care of operations and updates. Self-Hosted, you buy Hebby once and run it with our installer on your own server or VPS. Your data then never leaves your server.

Yes. Monthly plans can be cancelled monthly. You can export your memory at any time.

In a few minutes. Connect your agent, approve, get going. Hebby starts learning from the first session.