Memory

Automate

Memory holds the rules your agents reread before every job: your brand voice, a constraint, a name that matters.

Two scopes, and only two

A rule applies either to the whole business or to one named agent. The database accepts nothing else.

An agent rereads the business rules of the project, plus the rules that name it. It never reads another agent's rules. That isolation is enforced when the rules are gathered, not by the wording of a prompt.

The screen offers a third filter labelled for the project. It can never show anything, because no rule carries that scope.

Written by you, or inferred by MaShop

Two groups that never mix. The rules you wrote are framed with a solid border and count as approved instructions. The rules MaShop inferred from the work are framed with a dotted border, and the screen asks you to correct or delete them.

The distinction reaches the model. An inferred rule is handed to the agent flagged as inferred and unconfirmed, so it never treats a guess as an order you gave.

Write a rule

Add a rule opens three fields: the rule, what to remember, and who it applies to. The agent is picked from a list, never typed.

You can write a rule for an agent you have not hired yet. It applies the day you hire them.

Correcting turns an inference into your word

Correct reopens the same form, prefilled. Saving writes a new rule and removes the old one. There is no edit in place.

Correcting an inferred rule moves it into the group you wrote. Its dotted frame becomes solid, and agents stop reading it as a guess. That is the intent, and it is worth knowing before you click.

What actually reaches an agent

• 40 rules at most per run, the most recent first.
• Each value is cut at 600 characters, each label at 64.
• Anything past the fortieth rule is not sent at all.

Keep rules short and specific. A long paragraph competes with every other rule for the same room, and pushes older ones out of the run.

What an agent writes back

An agent records a fact only if you allowed it on its sheet, and that permission is off by default. Three facts at most per run.

Anything that looks like a secret is discarded twice, once as it is extracted and again before it reaches the table. That covers API keys, access tokens, private keys, card numbers, bank account numbers, and any rule labelled as a password.

Serving rules to one agent

The Memory block on an agent's sheet chooses which rules are served to it, and shows how many. Rules are written and removed on the Memory screen only.

An inferred rule cannot be adopted from that block. Rewrite it on the Memory screen to make it a standing instruction.

Deleting a rule

The confirmation says what happens: your agents stop applying it, and its text cannot be recovered. Nothing keeps a copy.

When a rule is merely wrong, correct it instead. The replacement then carries your wording, and the agents that read it stop guessing.

Who can change it

The owner and editors write, correct and delete rules. A viewer reads them. Writes are limited to 30 per minute, reads to 60.