Agents have no memory by default
A language model starts every session from zero. Early on, a startup's context lives in the founder's head and in old chats, so every session begins with re-explaining the product, the customers and last week's decisions. That works for one person. It falls apart with eleven agents.
The fix isn't a bigger prompt. It's a place where knowledge lives, gets updated, and can be checked.
Three layers
| Layer | What it holds | Who changes it |
|---|---|---|
| Sources | Raw material: call transcripts, CRM exports, tickets, contracts, research. Never edited after it lands. | Nobody. New sources get added, old ones stay as they were. |
| The brain | Plain-text pages about customers, competitors, products, people and decisions, with links between them. | Agents, through memory commits. |
| The charter | The rules: page formats, naming, what each agent may read and write, what needs a human. | People. Changes to the charter are always approved by a human. |
The pattern comes straight from open-source work. Andrej Karpathy's LLM Wiki idea file describes raw sources, an LLM-maintained wiki and a schema file. GBrain and ai-memory treat markdown as the source of truth and the search index as something you can rebuild. See the full credits.
Plain files, rebuildable index
The pages are markdown in a version-controlled repository. Search indexes and vector stores sit on top and can be rebuilt from the files at any time. That gives every change a history, every page an author, and the whole brain an export button. Humans and agents read the same files.
Brain first
Before an agent acts or answers, it checks the brain. Before it guesses, it asks.
When the gtm-chief writes to a prospect, it reads the account page first. When the cx-agent answers a billing question, it reads the pricing claim from the claim library, which lives inside the brain. Good answers get filed back as pages, so the next question starts further ahead.
One brain per company
Playbooks, skills and lessons are shared across the portfolio. Customer data is not. Each portfolio company has its own brain, and the hyena (cto) owns the access rules that keep them separate. A lesson like "this subject line works for CFOs" can travel. A customer's contract can't.
Building in a legacy industry?
We bring the agent crew and the distribution. You bring what you know about the industry.