How memory works

Memory is what makes the second conversation better than the first. It is deliberately built as something you can read, edit and delete — plain text, not a hidden model of you.

The two files used for conversation context

When memory is enabled, the main chat loads two short files as context. You will find them at Memory → Profile & rules. The browser extension and scheduled runs do not automatically load these files; their permitted tools can still retrieve relevant memory.

About you. Stable facts: your role, your licences, your fleet, your employer, your jurisdiction, how you want to be addressed. Things that stay true across conversations.

How you work. Working rules and house style: units, date format, citation style, terminology, how much detail you want, how you want drafts structured. Instructions to the assistant rather than facts about you.

Both are just text boxes. Open them, type, save. The change applies to your next message.

Keep these short. They are loaded into main-chat context, so they have a size budget. When they outgrow it, older and less-used facts are automatically summarised and moved into a longer note rather than being dropped — you keep the detail, the always-on files stay lean.

Do not put secrets here. No passwords, no API keys, no personal data you would not want in an AI conversation. There is nothing to gain by it.

Personal memory can inform other conversations on your account. A project prioritises relevant material but does not isolate it from other projects. Only save or reuse organisational material where you have authority to do so.

Longer material lives in the Brain

Anything too long or too specific for the two hot files becomes a Brain note — one note per project, decision, body of knowledge or recurring responsibility. Those are not read on every message; Metis fetches the relevant one when the conversation calls for it. See The Brain.

Your past conversations

Metis can also search your own earlier conversations, which is a separate thing from the curated notes above. Curated memory is what you decided is worth keeping; conversation recall is the raw history, searchable when you need to find what you actually said three weeks ago.

Long conversations do not lose their facts either. When a thread outgrows what the model can hold at once, the earlier turns are compacted into a summary that preserves identifiers, citations, decisions and root-cause statements word for word — and the full original transcript stays searchable if the exact wording matters.

How things get remembered

Three routes, and you are in the loop on all of them:

• You ask. "Remember that I always work in Cyprus time." It saves it directly, because you asked.
• It offers. When it notices something durable during a conversation, it proposes the edit and waits for you to accept.
• You correct it. Mark a reply with a thumbs-down and a note, and the correction becomes a staged proposal you can accept or reject.

Every update shows as a chip in the conversation, and you can undo it. Nothing is learned silently.

To make it forget something, tell it to — or open the file and delete the line. The second is faster and always works.

What it remembers about relevance

Memories are ranked by how relevant they are to what you are asking and how fresh they are. Older notes that are not marked as evergreen fade in influence rather than competing forever with what is true now.

Plans

Memory is a paid-plan feature. On the free plan Metis is stateless between conversations. On a paid plan you get the curated memory described here plus unlimited recall over your own conversation history. Downgrading removes access, not data. See pricing.