# Memory

> How the agent remembers facts between chats — the notebook it searches, the short notes and skills it learns by itself, how to ask it to remember or forget, and where to see and edit all of it.

Source: https://eggent.ai/docs/memory/

**Memory** is the agent's notebook. Unlike [context](https://eggent.ai/docs/context.md), it is not sent whole with every message: the agent searches it when needed, so it can hold lots of small facts.

## How it works {#how}

- The agent **saves** to memory what will be useful later: your decisions, preferences, agreements, facts it found.
- When a task may relate to something from before, the agent **searches** memory for it.
- The agent **deletes** outdated entries by itself or when you ask.

Memory lives in a `memory.md` file: the orchestrator has its own and so does every project. The agent only sees a project's memory while working in that project.

## What the agent learns by itself {#learned}

Besides the notebook, Eggent keeps two short lists that the agent **reads at the start of every chat**:

- **About you** — who you are and how you want to be answered: your work, language, tone, format.
- **About the work** — conventions and lessons that will matter again: where reports go, what not to do after you corrected it.

After some conversations — when you ask it to remember something, when you correct it, when it got past an error, or after a substantial piece of work, and now and then in a long chat — the agent looks back over the exchange in the background and decides whether anything is worth keeping. Most of the time nothing is. When it does save something, the note appears under its answer ("Noted about you: …"), and in Telegram as one short message. The look-back never delays an answer. It runs on the chat's own model — on Eggent AI while the workspace is on it; on your own model it stays there unless you [turn Eggent AI on for memory](https://eggent.ai/docs/models.md#services).

It can also write a **skill**: a reusable procedure for a kind of task it has just worked out. These skills are marked "Written by the agent" in **Settings → Skills**, are available in every project, and the agent changes only the ones it wrote itself — never one you made or installed. A skill that goes unused for a long time is marked, and much later moved to `skills/.archive/`. Nothing is deleted.

Safeguards:

- The lists are small (about 1,400 and 2,200 characters), so the agent has to choose what is worth a line.
- It never keeps passwords, keys or tokens, and refuses anything worded as an instruction to itself.
- It does not learn from light chats or from public project links.
- A note is a background fact, not a command: what you say in a chat always wins.

**Settings → Memory → What the agent has learned** shows everything: remove a note with the ✕, add one of your own, see the skills it wrote, or switch learning off (what is already saved keeps being used). The notes are in `learned.md` in the orchestrator's folder — plain text, editable by hand. To drop one in a chat, say "forget that I run a bakery".

## Asking it to remember or forget {#ask}

Just say:

- "Remember: I expect reports on Mondays by 10 am."
- "Remember that our main supplier is now Alpha, not Beta."
- "What do you remember about our suppliers?"
- "Forget everything about the old warehouse address."

## Viewing and editing {#edit}

**Settings → Memory.** In the top right, pick the orchestrator or a project. It is a plain text file: add, correct or delete entries by hand and save.

## Memory or context {#memory-or-context}

- **Context** — short rules needed in every message. The agent always receives them.
- **Learned notes** — a few short facts about you and the work, which the agent wrote itself and also receives every time.
- **Memory** — many facts, each needed now and then. The agent pulls them up on demand.

If the agent keeps forgetting some rule, move it from memory into the context.