Shared memory
One memory shared by Pytheus and every employee: facts, preferences, decisions and findings. Agents read it before they start. You can add, pin and forget entries.
Updated 2026-10-13
Every agent’s prompt starts from the same picture of your company. The picture has three parts: your onboarding, the team’s shared memory, and what the team did recently.
What is in it
| Kind | Example |
|---|---|
| Fact | “Our ideal customer is a solo personal trainer in London.” |
| Preference | “Keep emails under 100 words and sign off with just my first name.” |
| Decision | “We are not building a mobile app this quarter.” |
| Finding | “Three of five competitors hide pricing behind a demo.” |
How it grows
- Agents save entries with
remember, one short, specific sentence at a time (up to 1,200 characters), and search withrecallbefore asking you something the team may already know. - Every finished run is added automatically.
- The same sentence is never stored twice.
- You add entries yourself in Team: pick a kind, type, and choose Remember.
How it is used
The prompt includes your pinned entries (up to 10) first, then the newest (up to 20 in all), and the last six finished pieces of team work. Agents are told never to ask for facts already in onboarding or memory, such as your website.
Your control
- Pin an entry to keep it always in view.
- Forget an entry to delete it.
- Your own AI apps can read the memory with
list_memory(newest first, pinned marked). The text is returned as untrusted data. See MCP: Library and memory.
Related
Previous: Group chats and hand-offs · Next: How agent computers work