AI Assistant Memory & Objectives
Your Marmend AI assistant doesn't just answer questions — it learns how your team works and remembers it. This page explains what the assistant remembers, how to teach it, and how administrators can review and manage everything it knows.
Example: You ask the assistant to write an interview note, then correct it: "we always structure notes as Strengths / Concerns / Next step." The assistant saves that lesson — and every future note comes in your format, without being told again.
Two kinds of knowledge
| Objectives | Memories | |
|---|---|---|
| What they are | Standing goals that shape the assistant's behavior | Facts the assistant has learned about how your team works |
| Who writes them | You (administrators) | Mostly the assistant itself, or you |
| Example | "Reply to every candidate within one business day" | "Interview notes use the Strengths / Concerns / Next step format" |
Both are applied automatically on every conversation with the assistant — you don't need to repeat them.
How the assistant learns
There are two ways a memory gets created:
- Tell it directly. Say "remember that we sign all emails as 'The Team'" — the assistant saves it immediately and confirms.
- Correct it. When the assistant proposes an action (a note, a stage change) and you decline it with feedback, it can save the lesson so the same mistake doesn't happen twice.
Every memory shows where it came from: a badge marks whether the assistant learned it on its own (agent), you told it directly (user), or it was pre-configured (system). That way you always know why the assistant believes something.
Managing what the assistant knows
Administrators and owners can review everything at AI Assistant → Agent knowledge (the link appears in the AI assistant page header).
Objectives
- Add an objective (up to 20, each up to 500 characters) — short, clear goals work best.
- Reorder them — the assistant reads them in order of priority.
- Disable an objective temporarily without deleting it.
- Edit or delete at any time.
Memories
- Review every memory with its source badge, tags, and when it was last used.
- Edit a memory if the assistant learned something slightly wrong.
- Delete memories that are stale or incorrect — the assistant forgets them immediately.
- Add a memory manually if you want to teach the assistant something up front.
Good to know
- Knowledge is private to your workspace. Memories and objectives never leave Marmend and are never shared with external AI services or other companies.
- External AI assistants can't touch it. If you connect Claude or ChatGPT to Marmend, they cannot read or write your assistant's memory — it belongs to the in-app assistant only.
- There are limits by design. Up to 20 objectives and 100 memories per workspace keep the assistant focused. If the memory limit is reached, the assistant consolidates or asks you to prune.
- Deleting is safe. If you remove a memory the team still relies on, just tell the assistant again — it will re-learn it.