Estratos

How memory behaviour works

Three sets of instructions decide how your assistant handles memory. Two of them come from Estratos. The third is whatever you have typed into your assistant's own settings, which has nothing to do with Estratos but shapes what it does anyway.

What your assistant does by default

  1. Before anything else, your assistant loads your hierarchy from Estratos: the layers, what they are called, and how you name things. That is how it knows where your memories live and what they mean. Once per conversation.
  2. To answer a question, your assistant matches your words to an Estratos folder (see below), then reads everything that applies from the top of the tree downwards. Where two memories disagree, the more specific one wins. If the folder has nothing recorded, your assistant says so rather than inventing an answer.
  3. When you ask your assistant to remember something, it writes the title itself and keeps your own words alongside the memory, so whoever reviews the memory can see what prompted it. If a similar memory already exists, your assistant updates that one rather than adding a second copy.
  4. When you have not asked, your assistant suggests and waits. At the end of a task it looks back over the conversation for a decision, a correction, a lasting fact about a client, a convention your team follows, or a trick it worked out itself. You get two or three suggestions at most, and nothing is saved until you agree.

That last kind, the trick your assistant worked out for itself, matters most and is the easiest to lose. Nobody thinks to ask for it, because nobody else knows it.

How your assistant picks the Estratos folder

Your assistant reads the names out of your sentence and matches them to your Estratos layers, one name per layer. Your layers' example values tell it how you spell things, so "the Pennine Joinery VAT work" becomes a client and a service.

Two things follow from this. Name the client or the job plainly and your assistant will find the folder. And if it keeps guessing wrong, the fix is not a longer instruction: add real example values to the layer, so your assistant learns how you spell things.

Changing what your assistant does

To change Do this Who can
Stop it suggesting saves Set Settings › Organisation › Agents to Off. Asking it directly still works. Owners, Admins
Where it files things Add real example values to the layer, in your hierarchy Owners, Admins
A rule everyone follows Save the rule itself as a memory Anyone

Saving the rule as a memory is the useful habit, and it needs no special syntax. Say what the rule applies to, and your assistant works out where it belongs:

Remember this as a rule for all VAT work, for every client: every VAT
deliverable must quote the relevant HMRC notice number, and must be checked
by a second person before it goes to the client.

"For all VAT work, for every client" is what makes that rule crosscutting. The rule does not belong in one client's folder, so your assistant attaches it to your Services layer at VAT instead. Estratos calls that a tagged memory, and that is why the rule reaches every colleague on every assistant without anyone needing to know the rule exists. A fact about one client only is a scoped memory, and lives in that client's folder. Scoped vs. tagged covers the choice.

What survives clearing a chat

Only what was saved to Estratos. Everything else lived in the conversation, and clearing the conversation throws all of it away.

A suggestion you left unanswered is gone, because it was only ever part of a conversation. Something already waiting for an approver is different, and stays. And a long conversation is not memory: if something important came up an hour ago and was never saved, clearing loses it.

Before you clear a chat that produced something worth keeping, ask: anything here we should save?

Next steps