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

# Context and memory

> Teach Fella your vocabulary and understand what it remembers about a workspace.

Fella can use context you write and vocabulary it learns from explicit corrections. Context can guide interpretation, but it is not evidence and cannot grant a tool or permission.

## `fella.md`

<Steps>
  <Step title="Open the context editor">
    Type `/context`, or create `fella.md` at the workspace root.
  </Step>

  <Step title="Describe stable conventions">
    ```markdown filename="fella.md" theme={null}
    Statements are one file per month, named statement-YYYY-MM.csv.
    Amounts are in GBP. "Spending" excludes transfers between my own accounts.
    The category called housing is part of rent for my reports.
    ```
  </Step>

  <Step title="Ask with the new context">
    Fella reads the file at the start of each question. It can use the conventions to interpret a request, but figures still need tool evidence from workspace data.
  </Step>
</Steps>

## Folder memory

Folder memory lives as human-readable Markdown in Fella's application data directory, separate from the workspace. It is keyed to the folder path and re-read for each question.

```text filename="fella commands" theme={null}
/memory
/memory forget
```

Automatic learning is deliberately narrow: an ordinary question teaches memory nothing. A follow-up recognized as a correction can add, update, or deduplicate a vocabulary note, using the active model to reconcile it with existing notes. Fella also records a bounded local episode log, but does not read that log back as memory or cache old query results as facts.

<Warning>
  A correction may trigger an additional model request to reconcile vocabulary. With a hosted provider, that correction and the existing vocabulary notes are sent to the provider.
</Warning>

## Context is not permission

`fella.md` and memory can influence how the model interprets a question, but
they cannot grant it a new tool or a write capability. The read-only boundary
comes from the registry and engine.

## When to use which

| Need                                                                     | Use                                 |
| ------------------------------------------------------------------------ | ----------------------------------- |
| Stable file conventions                                                  | `fella.md`                          |
| An explicit vocabulary correction that should carry across conversations | Folder memory                       |
| Provider-specific or domain-wide guidance                                | `fella.md` plus the normal question |
| New external data                                                        | Mount the files in the workspace    |

<CardGroup cols={2}>
  <Card title="Experimental MCP" icon="puzzle" href="/extensions-and-mcp">
    Read the inert extension boundary for future forks.
  </Card>

  <Card title="The model prompt" icon="scroll-text" href="/concepts/harness">
    See where context enters the reasoning loop.
  </Card>
</CardGroup>