Join the waitlist to take part in the alpha.
Let the agents work

MEMORY

A memory you can open and correct.

What your agents remember is not buried in a model: it is a record in your workspace, readable, editable, deletable.

A memory is a record

Memory is not a hidden store somewhere under the conversation. It is an object in your workspace, with its fields: a name, a one-line description, the content, and a kind.

Four kinds, and no more: a fact, a preference, a piece of feedback, a reference. So you always know what sort of statement you are looking at.

A memory attaches to what it is about: a client, a contract, another memory. It does not float in a vacuum, it hangs off your business.

Nothing is memorised behind your back

An agent does not log the conversation as it goes. Writing a memory is an explicit act, one it decides on and one you can find again.

Before writing, it checks. If a close memory already exists, the write is refused and the agent is sent back to the existing record, to correct it rather than duplicate it. A memory that fills up with variants of the same fact is a memory that will soon lie.

The index always, the text on demand

The agent permanently carries the index of your memories: forty entries at most, a name and a line each. It is short on purpose, and that is what makes it carryable at all times.

It opens the full content only when it needs to, or searches by meaning when it does not know which one to open.

Your personal memories come before workspace-wide ones in that index, and a memory with no owner counts for everyone.

Long conversations do not break

This is a separate mechanism, and worth separating: compacting a conversation is not memorising. When an exchange goes past its budget, everything before your last message is summarised, and the tail is kept word for word.

If the summary fails, the session does not die. It falls back to truncation, and the agent tells you rather than carrying on as if nothing happened.

How it works

You correct an agent: for this client we bill a fixed fee, never by the day.

  1. It offers to remember: a name, a one-line description, the content, and the kind. Here a preference, not a fact.

  2. It checks it is not duplicating. A search goes out over nearby memories, and if one exists the write gives way to a correction.

  3. It hangs the memory off the client, so it will be found from the file and not only from a search.

  4. The line enters its index. From the next conversation on, on a different subject, it knows the preference exists without loading it.

  5. You keep control. The record opens, reads, corrects and deletes like any other record in your workspace.