> ## Documentation Index
> Fetch the complete documentation index at: https://fal.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Memory and documents

> How project memory captures decisions and preferences, and how uploaded documents become project context.

Project memory lets the agent recall what happened in earlier chats. It has two parts: notes distilled from the chats, and documents you upload.

## Memory notes

After each turn in a project chat, fal Agent ingests the turn into the project's memory. A background pass distills notes and a short primer that summarizes the project. Notes have a kind:

| Kind           | Example                                              |
| :------------- | :--------------------------------------------------- |
| **Decision**   | "Final palette is teal and rust."                    |
| **Preference** | "Prefers 4:5 for all product shots."                 |
| **Style**      | "35mm, shallow depth of field, warm grade."          |
| **Entity**     | "Mara is the lead character. Red coat, silver hair." |
| **Fact**       | "The launch date is March 12."                       |

Notes are short, standalone statements. A new note that repeats an old one replaces it.

### Manage notes

Open the rail's **Context** tab to see the notes. You can pin a note so it is always in context, edit its text, change its kind, or delete it. You can also ask the agent to remember something directly:

```text theme={null}
Remember: every video ends on the logo card, 1.5 seconds, no audio.
```

### How the agent uses memory

At the start of a turn the agent receives the primer and the pinned notes. It can search the rest of the memory, including verbatim history from earlier chats. It can also search the project's media by text or by image similarity. It saves a new note when a decision is made.

## Documents

Upload PDF, DOCX, TXT, Markdown, CSV, TSV, or JSON files with **Add resource → Upload documents**. You can also drop them on the project page. Documents become project context.

* The text is split into chunks and indexed for search. A document is indexed up to about 250,000 characters.
* A one-line summary of each document appears in the document roster. The summary says what the document contains and what it is useful for.
* The agent reads a document in pages when it needs the full text.

Documents uploaded to a project are indexed far more deeply than a document attached to one message, which is read only in part. Put briefs, scripts, and brand guides in the project.

A document's status shows **Ready**, **No text** when nothing could be extracted, or **Failed**. Failed documents can be retried.

## Memory and retention

Derived memory and indexed documents have no expiry. Both are turned off when a finite [retention policy](/docs/documentation/agent/retention) applies to the workspace.