Skills and memory

This guide follows the public 2026.10.19 release. Use skills for reusable instructions, profile documents for agent-specific guidance, and Brains for explicit saved memories.

Create a text skill

1. Open Settings → Skills → New text skill.
2. Fill in ID, Name, Description, and Content. Alternatively, click Choose text file to load a .md, .markdown, or .txt file into the form for review.
3. Choose a Lifecycle: available, experimental, deprecated, or disabled.
4. Check Assign globally if every agent should have access, or leave it unchecked for individual assignment.
5. Click Create skill. Loading a file into the editor does not save it until this step.

To assign a saved skill, open Installed skills, select it, and use its Assignments section. You can also open Settings → Agents → [agent] → Skills and check individual assignments. Assignments refer to the same shared skill; they do not create separate copies. A global skill cannot be individually unchecked in the agent view.

Use Installed skills to search by name, ID, or description, filter by lifecycle/assignment, edit operator-managed instructions, or delete a skill. Deletion removes the shared copy and its assignments.

Source: text creation and assignments

Import a repository skill

1. Open Settings → Skills → Import from repository.
2. Enter the repository and Skill path. Select a directory for a package with companion files, or SKILL.md for instructions only.
3. Under Ref & automatic updates, choose a ref if needed. HEAD uses the repository default; tags and commit IDs stay pinned. Automatic sync is off unless enabled.
4. Click Preview skill. Read the actual instructions, included files, and previewed revision.
5. Review Install ID & availability, then click Install skill.
6. Assign it to the intended agents if you did not assign it globally.

The interface installs files; it does not execute package files during import. If the repository changes after preview, preview again before installing. If an install response times out, reload the library to check whether it completed before retrying.

Source: repository preview and import

Update an imported skill

Select it under Installed skills. Its Source section controls the repository, ref, path, and automatic sync schedule. Save source saves that configuration; Sync now fetches the upstream content. Refresh status checks an ongoing or uncertain result. Schedules use the operator timezone.

Upstream-managed instructions are read-only in the editor; lifecycle and assignments remain editable. Disabling automatic sync retains the installed content. A text-only source can be removed while retaining its text and assignments. A package with companion files cannot be detached here without copying; disable automatic sync instead. Private HTTPS credentials are not configured in this skill form.

Source: source controls and limitations

What assigning a skill does

Assignment makes the skill available for discovery. It does not inject the full instructions into every prompt or guarantee that a particular turn uses them. The agent can list skills and explicitly load the relevant one. The planner does not automatically select full skill bodies. A disabled skill is unavailable; experimental and deprecated are lifecycle labels, not equivalent to disabled.

Source: explicit loading · Source: lifecycle availability · Source: skill discovery and load

Edit an agent's profile documents

Open Settings → Agents, choose the agent, then Profile documents. Select SOUL.md, RULES.md, ORIENTATION.md, PREFERENCES.md, TOOLS.md, or DreamMemory.md. Edit the Markdown and click Save profile documents.

These are agent guidance documents. Writing a permission into a document does not create an operating-system sandbox or an MCP grant; configure actual access separately in Tools and mods.

Source: profile document editor

Create and correct saved memories

1. Open Brains and select the Agent owner.
2. Click New memory.
3. Add a Title, Content, and optional Source references, one per line.
4. Click Create memory.

Search within the selected agent's memories, then select one to read or edit it. Click Save memory to save an edit. If another change caused a revision conflict, use Reload memory and review the latest version before saving again.

Memories you create or edit through the operator interface are operator-owned; agents cannot modify operator-owned records. Memories are separated by agent owner. They have no automatic expiration or full-memory preload: the agent uses explicit memory tools to retrieve them. A saved fact and its references may become stale; neither proves current runtime state.

Source: Brains UI · Source: ownership and revisions · Source: operator mutations · Source: retrieval contract

Check the selection before deleting

Delete memory and Delete selected memories act directly in this release; there is no confirmation dialog in the Brains page. Select page selects only the visible page. Bulk deletion runs independently for each record, so some deletions may succeed while others fail. Read the per-record result before retrying. This page has no restore button.

Source: deletion behavior

Optional embeddings

Open Settings → General → Brain memory to configure embeddings globally. This release supports Google and Ollama embedding adapters; a working chat connection to another provider does not establish embedding support. Choose a supported provider and embedding model, test it, then enable and save the settings. A manually entered model ID requires a successful test before saving an enabled configuration. Reindex all memories asynchronously rebuilds the index in the background; the page reports indexed, pending, and failed counts. Partial coverage remains searchable lexically, and saved-memory editing remains separate from indexing.

Enabling a remote embedding provider involves sending memory text to that provider for indexing. Choose the provider with the sensitivity of your saved content in mind.

Source: embedding controls · Source: supported providers and text requests

Scheduled dreaming

Open Settings → Agents → [agent] → Dreams. Enable scheduled dreaming, set the cron schedule, choose a model override or inherit the chat model, edit the prompt, and click Save dream settings. The schedule follows the operator timezone. Dream on quiet days controls whether a day with no new material still generates a diary from retained continuity; when off, such days skip model work.

Check Recent activity for completed, partial, failed, or interrupted cycles and any phase errors. Browse retained dream entries in Archive → Dreams. Dream continuity and explicit Brains records are separate features.

Source: dream settings and activity