Skip to main content

Creating a memory

Setting up a memory is like briefing a new assistant on their first day: you tell them what this notebook is for, and they take it from there. You can either talk it through and have AI write the first pages, or open an empty notebook and fill it yourself.

Both start on the Memories page.


Before you start

Have a rough answer to one question: when should an agent open this memory? That sentence is the single most important thing you'll write — it's what every agent reads to decide whether your notes are relevant to what it's doing.


Option 1 — Create with AI

Best when you have a lot in your head and don't want to type it all into forms.

  • Click Create with AI and describe what you want remembered — a client, your team, a product, how you like things done.
  • It asks a few follow-up questions, then drafts the whole memory: a name, the "when to use" description, and a set of observations with the people and things already picked out.
  • You land on a review step before anything is saved.

On the review step you can toggle any observation off, edit its title, wording, category or tags, rename the memory, and attach reference files. Nothing is committed until you press Looks good — commit to memory.

It can also add to a memory you already have

On the review step, switch the destination from Create new base to an existing memory. Use this when the facts belong with something you've already built rather than in a new store.


Option 2 — Start empty

Best when you know exactly what you want and would rather fill it yourself, or you just want somewhere for an agent to start writing.

Click New memory and fill in three things:

FieldWhat to put
NameShort and recognizable — Clients, Engineering knowledge, Personal
When to use this memory baseRequired. The sentence agents read to decide whether to consult it
VisibilityPersonal (only you) or Organization (everyone on your team)

You can also give it an emoji, an icon or an image so you recognize it at a glance.

Writing a good "when to use"

Weak descriptions get ignored by agents. Strong ones say what's inside and when to reach for it:

  • ✅ "My consulting clients — scope, rates, contacts and where each engagement stands. Consult before writing to a client or quoting work."
  • ✅ "My writing voice and style rules. Use before drafting anything in my name."
  • ❌ "Client stuff"

If you're not sure, use the Start from a template dropdown — it fills the description with a ready-made version you can edit. There are templates for personal work, engineering knowledge, projects, personal assistance, sales, support, habits and goals, research, writing voice, hiring, clients and meetings.


Adding facts yourself

Most observations get written by your agents while they work. To add one by hand, open the memory, go to Observations, and click New observation.

Acme renewed for 12 monthsClients

Acme Corp signed a 12-month renewal in Q3 at the standard rate. Sarah Chen owns the account and asked to be the only point of contact for billing.

Tags
renewalq3
Entities
Acme CorporganizationSarah ChenpersonQ3 renewalproject
Relationships
Sarah ChenWORKS_ATAcme CorpSarah ChenOWNSQ3 renewal

You write the title and the fact. The people and things in it — and how they connect — are picked out for you.

Keep each note to one fact, written so it still makes sense a year from now. "Acme renewed" is a good note; a pasted email thread is not.


Troubleshooting

What you seeWhyWhat to do
Agent ignores the memoryIts description doesn't match what the agent is doingRewrite "when to use" to name the situations, not just the contents
Agent can look things up but never records anythingIt's connected as Read onlyChange its access to Read & write on the agent
A fact is wrong and the agent won't fix itCorrecting needs Full accessRaise its access, or edit the observation yourself
The same person appears twice under slightly different namesTwo notes named them differentlyRun Consolidate, or merge them from the memory's AI assistant
You can't change the memory's language model settingsMemories are built to work one way and that setting is fixed at creationNothing to do — this is intentional, so old and new notes stay searchable together

What's next