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

# Memory

> Persistent memory system with keyword and semantic search

Bolly remembers everything across sessions. The memory system uses a file-based library with two search methods: BM25 keyword search (always available) and vector semantic search (requires Google AI key).

## How it works

Every memory is stored as a file in `~/.bolly/instances/{slug}/memory/`. Bolly creates, reads, and manages memories automatically during conversations — you don't need to tell it to remember things.

When you mention something important — a goal, a preference, a project detail — Bolly writes it to memory. Next time it's relevant, Bolly recalls it.

## Memory categories

Bolly organizes memories into folders:

| Folder         | Purpose                                                   |
| -------------- | --------------------------------------------------------- |
| `about/`       | Facts about you — name, role, interests, background       |
| `preferences/` | Your habits, likes, dislikes, work style                  |
| `moments/`     | Shared experiences and conversations worth remembering    |
| `pinned/`      | Always-loaded memories (injected into every conversation) |

You can also create custom folders. Bolly will use whatever structure makes sense for the conversation.

## Search methods

### BM25 (keyword search)

Always available, no extra setup needed. Works well for finding memories with specific terms.

### Vector search (semantic)

Requires a [Google AI key](/docs/installation/api-keys). Uses Gemini embeddings to find memories by meaning, not just keywords. For example, searching "feeling stressed about work" can find a memory about "deadline pressure on the project" even without matching words.

<Info>Vector search uses LanceDB (embedded, no external server needed) and Gemini 2 embeddings at 768 dimensions.</Info>

## Pinned memories

Any memory file placed in the `pinned/` folder is loaded into Bolly's context for every conversation. Use this for things Bolly should always know — your name, key goals, important boundaries.

## Media memories

Bolly can store more than text. Supported formats:

* **Images:** JPG, PNG, GIF, WebP
* **Documents:** PDF
* **Video:** MP4, MOV, WebM
* **Audio:** MP3, WAV, OGG

Media files are embedded into the vector store (with Google AI key) for semantic search alongside text memories.

## Memory tools

Bolly has 5 memory tools it uses automatically:

| Tool            | What it does                              |
| --------------- | ----------------------------------------- |
| `memory_write`  | Create or update a memory file            |
| `memory_read`   | Read a specific memory                    |
| `memory_search` | Find relevant memories by query           |
| `memory_list`   | Browse memories in a folder               |
| `memory_forget` | Delete a memory that's no longer relevant |

## Frontmatter timestamps

Each memory file automatically gets timestamps:

```markdown theme={null}
---
created: 2025-03-28
updated: 2025-03-30
---

User is working on a SvelteKit project called Bolly...
```

These help Bolly understand how recent a memory is.

## Managing memories manually

You can directly edit files in `~/.bolly/instances/{slug}/memory/` with any text editor. Changes are picked up on the next conversation or context refresh.

To force a full rescan and reindex of the memory library, use the reindex endpoint or restart Bolly.