# Memory

Volae has one memory layer. The coworker, the chat assistant, reply and outreach drafts, GEO comparisons and your own agents over MCP all read and write the same place.

## What it holds

| Scope | What | Example |
|---|---|---|
| `user` | Your preferences and standing instructions | "No exclamation marks in email" |
| `product` | Positioning, features, prices, rivals | "Refund window is 14 days" |
| `person`, `company` | Facts, promises, what matters in the relationship | "Promised an answer on the refund by the 15th" |
| `theme` | How a theme got here and where it landed | "Team plans: next quarter, 15 people waiting" |
| `goal` | The coworker's working notes on a goal | "Comparison page live, mention rate up 4 points" |
| `coworker` | A summary of each run | "Handled 3 negative reviews, replied to 2" |

Things you can read directly from connected data (plan, MRR, last active) aren't memories. Read them with `get_person_context`.

## Reading

| Action | What it returns | Unconfirmed memories |
|---|---|---|
| `recall` | A search by subject, optionally with text. Searches words and meaning | Left out, unless you pass `includePending` |
| `get_context_pack` | What to read before working on a person, company, product, theme or goal, including your standing instructions and preferences | Included. Check each item's `status` and `tainted` |
| `list_memories` | Everything stored, by scope and `status`, newest first | Included. Use this to read back what you just wrote |

Use `get_context_pack` before you act and `recall` when you need something specific.

## Writing

`remember` writes one memory with where it came from, and returns it with its `status`. What you get depends on who wrote it:

| Written by | Status | `tainted` |
|---|---|---|
| You, in the web app | `confirmed` | No |
| An agent, with `confidence` of 0.8 or more and a `source` other than `user` | `confirmed` | Yes, until you confirm it |
| An agent, any other way: no `confidence`, under 0.8, or `source: "user"` | `pending` | Yes, until you confirm it |

An agent saying you told it something doesn't make it so: that stays `pending` until you confirm it. Since `recall` leaves pending memories out, an agent that writes one and recalls it won't find it. Read it back with `list_memories`, and tell the user it's waiting for them.

- A new memory on the same subject and key replaces the old one and keeps the history. If it isn't clear which is right, both are flagged as a conflict and it lands in "needs you".
- Anything from an email, page or post is stored as a fact and marked `tainted`, and so is anything an agent writes, until you confirm it. It stays data, never instructions.
- **Standing instructions can only be written by you, in the web app.** No agent, email or page can write one.
- Agents can't change what you wrote. If one disagrees, it writes a new memory on the same key, and the conflict comes to you. Only you can settle a conflict.
- Only you can allow a memory to be used in customer replies.

## Seeing and changing it

**Settings → Memory** shows everything by scope. You can edit, pause, forget and restore, and every change keeps a version. Agents can list memories with `list_memories` and change their own with `update_memory`. Deleting a person deletes their memories.

## What your visitors can't reach

The chat assistant on your site doesn't search memory. It answers from your public, confirmed knowledge, a few product facts (one-liner, audience, tone, and pricing and policies only if you wrote them yourself) and the signed-in customer's own account, so nothing internal can come out in a chat.
