# Knowledge and memory

{/* AUTO-SYNCED SOURCE: this page lives in apps/app/src/modules/agent/docs/ and is mirrored into the docs app by `bun sync:module-docs`. Edit it in the module, not in apps/docs. */}

<Lead>
The assistant does not forget everything after each chat: the knowledge vault keeps your workspace's knowledge, personal memory keeps your individual preferences. Both stay under your control, with clear ways to manage and forget.
</Lead>

## The workspace knowledge vault

The **Wissensspeicher** (knowledge vault) at `/agent/memory` (reachable via the link in the footer of the [sidebar](/tools/agent/chat-und-projekte#the-sidebar)) is your workspace's knowledge base. It works semantically: it searches by meaning, not just exact words. "Commission for dual agency" will also find a note that is phrased differently.

- **Search:** On the vault page you search all entries directly. The assistant itself also draws on it in chats whenever it fits the question.
- **Store:** You can deposit knowledge deliberately, such as office standards, phrasing templates, or property knowledge.
- **Silent learning:** The vault additionally learns silently from conversations: recurring, useful facts from chats land as entries automatically, so knowledge does not disappear into old chats.

At the top of the file tree on the page itself there is an **Anleitung** (guide) button. It holds the same information as below, only shorter and right next to the editor. On your first visit it opens by itself.

### Where does what belong

These are plain text files. There is no format to satisfy and no table to fill in. Full sentences work best, the way you would explain something to a new colleague on their first day. When in doubt, use `company/`: something in roughly the right place beats nothing at all.

| Folder | What belongs in it | Who writes |
|---|---|---|
| `company/` | Who you are, how you write, how you work. Company profile and team keep themselves in sync from your settings; your own notes below stay untouched. | you |
| `markets/` | Which towns and price bands you work in, what happens in which season. | you |
| `decisions/` | What applies at your office, with a date and a reason. The assistant stops contradicting these afterwards. | you |
| `learned/` | Observations from conversations. You may read along, correct, and delete. | the assistant |
| `projects/` | Knowledge about one specific property or mandate that should apply permanently. | both |
| `documents/` | The text from uploaded PDFs and text files. | both |

Two core files are protected and cannot be deleted: `ASSISTANT.md` (the assistant's ground rules) and `INDEX.md` (the table of contents, which maintains itself).

### What you can safely do

Write, rewrite, reorganise, delete: none of it can break a booking, a property, or a file. It only changes what the assistant knows. Delete a file and it no longer knows that content, which is the only consequence. And because search goes by meaning rather than folder name, the wrong folder costs you overview, not function.

Two things still do not belong here: passwords and credentials (those live in the credential vault) and notes about individual people that have no business being there. Everyone in the workspace can read along.

### Will this get out of hand over the years?

No. For every answer the assistant reads only the two small core files in full; everything else it searches by meaning and pulls in the matching passages. Whether the vault holds 12 or 400 files makes its answers neither slower nor more expensive.

For your own overview, three habits help:

- **One topic per file.** Once a file becomes a catch-all, split it up.
- **Overwrite what is outdated instead of appending.** The assistant does not weigh by date, it believes both versions. Contradictions are the only real problem in a vault that has grown.
- **Look into `learned/` once a quarter** and throw out what does not fit.

The current size of the vault is shown at the bottom of the file tree. A well-kept vault sits at one to two megabytes after three years.

## Personal memory

Personal memory belongs to you alone: colleagues have their own, nobody sees anyone else's entries.

- **Remember in chat:** Just tell the assistant "Remember that I always write listings in formal address". It proposes the entry, you confirm with **Zustimmen** (agree) or decline.
- **Forget in chat:** Just as direct: "Forget that ..." removes an entry again.
- **Manage:** Under **Account > Assistant** (Konto > Assistent) in the **Gedächtnis** (memory) section you see all entries, can add new ones, delete individual ones, or **forget everything** with one click.

## What is the difference?

<Comparison columns="Knowledge vault,Personal memory">
  <CompareRow label="Who owns it?" values="The whole workspace; everyone benefits,Only you; nobody else sees it" />
  <CompareRow label="What belongs in it?" values="Office facts and standards (commission rules; text blocks; property knowledge),Your preferences (tone; signature; way of working)" />
  <CompareRow label="Where to manage it?" values="On the /agent/memory page,Under Account > Assistant in the memory section" />
</Comparison>

## Frequently asked questions

<Faq>
  <FaqItem q="Can colleagues see my personal memory?">
    No. Personal memory entries are bound to you. Every person in the workspace has their own memory.
  </FaqItem>
  <FaqItem q="Can I delete everything at once?">
    Yes. Under **Account > Assistant** in the **Gedächtnis** (memory) section there is an action to forget all personal entries at once. You delete individual entries there as well.
  </FaqItem>
  <FaqItem q="Where do entries come from that I never created?">
    The knowledge vault learns silently from conversations and stores useful facts automatically. You can inspect and remove every entry on the vault page.
  </FaqItem>
</Faq>

## Further reading

How the assistant uses this knowledge in answers is shown in the [capabilities reference](/tools/agent/faehigkeiten). All personal toggles are covered in [Settings](/tools/agent/einstellungen). Back to the [overview](/tools/agent).