What Leo remembers
Overview
Memory is Leo's long-term memory of you, kept between sessions: stable facts, preferences and how recurring things are usually done. Its purpose is simple — you should not have to re-explain what you already said.
Memory is not the knowledge base and not the conversation thread. The knowledge base is content you deliberately put in and can browse; a thread is one conversation and ends with it. Memory is what the assistant retained about you.
When to use it
- You keep repeating the same context to the assistant at the start of every conversation.
- You want the assistant to follow a preference — a format, a tone, a way of naming things — without being told each time.
- You want to check, or remove, something the assistant remembers.
Step-by-step
- Tell Leo what matters. "Remember that I run rollout calls on Mondays", "I prefer short answers".
- See what it kept. Ask Leo to list what it remembers about you.
- Forget an item you no longer want kept. Forgetting is physical, not a flag — the record is removed.
- Ask for a deferred action when you need one later: "remind me in an hour", "send this next week". These are stored as a record of what you asked for and executed on time.
Memory, Knowledge and Context
| What it is | Who put it there | |
|---|---|---|
| Memory | Stable facts and preferences about you | The assistant, from your conversations |
| Knowledge base | Notes and documents that answers draw on | You, deliberately |
| Context | Everything assembled for one answer | Computed per answer, not stored |
Common mistakes
- Expecting memory to hold documents. Material to search belongs in the knowledge base; memory holds facts about you.
- Expecting a thread to persist as memory. A conversation ends; what was worth keeping is kept separately.
- Assuming an external client can erase it. Forgetting a memory is never exposed over MCP — there is no confirmation screen there, and the removal is not recoverable.
Best practices
- Say what should be remembered explicitly; the assistant is deliberately conservative about inferring it.
- Review the list occasionally — a preference that was true in spring may not be true now.
- Keep sensitive material out of memory and use no-logs mode for conversations that should leave no record.