Embeddings and persistent memory

What persistent memory does#

By default, your agent's context is limited to the current conversation. Persistent memory lets it remember information across sessions, recall past work, and build on previous interactions.

How it works#

Le Bureau uses an embedded vector database to store and search your agent's memories. When enabled:

1. The agent creates embeddings (numerical representations) of important information during conversations.
2. These embeddings are stored on the desktop's persistent disk.
3. When you start a new conversation, the agent searches its memory for relevant context.
4. Retrieved memories are injected into the conversation, giving the agent continuity.

What you need#

Persistent memory requires an embeddings API key and model, separate from your main AI provider key. Embeddings use a specialized model optimized for converting text into vectors.

Supported embeddings providers#

Setting up persistent memory#

During desktop creation#

1. Click New Desktop.
2. Configure your main AI provider (Anthropic, OpenAI, or OpenRouter).
3. Expand the Embeddings section.
4. Select an embeddings provider (OpenAI or OpenRouter).
5. Enter your embeddings API key.
6. Choose an embeddings model.
7. Create the desktop.

On an existing desktop#

1. Open your desktop's settings.
2. Add or update the embeddings configuration.
3. Restart the desktop.

What memory looks like in practice#

With memory active, your agent can:

• Remember project context: "We are building a REST API with Express and PostgreSQL."
• Recall past decisions: "Last time, we decided to use JWT for authentication."
• Track progress: "The user asked me to set up three microservices. Two are done."
• Learn preferences: "The user prefers TypeScript over JavaScript."

Without memory, every new conversation starts from scratch.

Where memories are stored#

Memories live in ~/.openclaw/workspace/memory/ on the desktop's disk. As long as the desktop is not destroyed, they persist across reboots, stops, and restarts.

You can inspect the memory directory from the terminal:

ls ~/.openclaw/workspace/memory/

Cost#

Embeddings API calls are charged by your provider based on tokens processed. The cost is typically very low, a few cents per million tokens for text-embedding-3-small.

Tips#

• You can use a different provider for embeddings than for your main model. For example, Anthropic for chat and OpenAI for embeddings.
• If you do not need memory, skip the embeddings configuration. Your agent still works, it just will not remember past sessions.
• Memory quality improves over time as the agent accumulates context about your projects and preferences.