Instructions, todos, notes

Besides events and facts, mnemos holds four more things. All are created in words and live apart from retrieval by meaning.

Instructions

A procedure the assistant follows when the case comes up.

— Remember how I want a supplier checked: registration details against the registry first, then three years of litigation, then who the owners are. Report as a table, no padding.

After that "check this supplier" is enough — the instruction finds itself.

Name an instruction after the question it answers. "How to check a supplier" is found better than "Counterparty verification": the assistant searches by how you ask.

This is the one layer with a floor: an instruction that does not match does not come back at all. An event or a fact can be read and discarded; instructions are followed, so an empty answer beats the nearest thing lying around.

Do not turn an instruction into a log. Appending to the end accumulates incidental words, and the instruction starts surfacing on unrelated questions. Put a change where it belongs: "in the supplier check, replace the litigation step with this".

Schedules do not belong in an instruction: it says how, and when is the scheduler's business.

Todos

— Add a todo: get their confirmation of the dates by Friday. — What is outstanding on Antarius?

Todos never fade with time — the forgotten obligation is the one that matters most. They are retrieved by date and state rather than by meaning, so they do not turn up inside an ordinary answer: ask for them.

They are closed as done or as dropped, and either way they stay. "We considered this and said no" is the answer you want next time.

Pinned notes

A short rule mixed into every answer in its area.

— Pin in work/antarius: prices are discussed with Ivanov only, nobody else is authorised.

A pin applies to its area and everything under it. It is a place for a few permanent things: a rule, a restriction, somebody's prohibition. Everything else is a fact.

Handing a conversation to another model

— Save the context of this conversation as "audit" so I can continue in another client.

What is written is a briefing: what was concluded and what is still open. In the other client:

— Load the "audit" context.

A name already taken is not overwritten: it is saved under a neighbouring one and the reply says which. Contexts live a month and are extended on each save — this hands a conversation over rather than storing a transcript.

Bookmarks for recurring jobs

When the assistant runs something on a schedule it leaves a bookmark: what it did last time and what the next run should do. It is kept apart from memory and never appears in answers to questions.

You see it when you need to know where the last run stopped:

— Where did yesterday's registry check stop?