Configure agent tools, skills, and memory

Models decide what to say. Tools determine what an agent can do. Skills provide reusable instructions. Memory carries selected knowledge across sessions.

Grant a tool

1. Edit the agent.
2. Open Tools.
3. Enable a routine tool or expand Advanced tool access.
4. For MCP, select only a connected server and review its exposed tool names.
5. Save.
6. In chat, ask for a harmless operation that requires the tool.
7. Inspect the tool call and result.

If you later set a pipeline-step or schedule tool override, that explicit list replaces the agent’s normal tools for that run.

Choose assigned skills versus Skill Loader

• Assigned skill: injected every time in chat, pipelines, and schedules. Use for core role procedure.
• Skill Loader tool: lets the agent discover and load instructions mid-conversation. Use for occasional specialist work.

Too many assigned skills consume context and can introduce conflicting procedures.

Establish memory

Enable Persistent Memory, save the agent, then either:

• tell the agent what to remember and let it call save_memory; or
• edit Agent → Memory directly.

The memory tool replaces the current MEMORY.md, so an agent saving memory must include everything it intends to retain.

Good memory:

## Operator preferences
- Use concise status updates.
- Never publish without explicit approval.

## Project conventions
- User-facing changes require a desktop and mobile check.

Poor memory includes API keys, full transcripts, one-time payloads, speculative facts, or copied reference manuals.

Audit context pressure

If an agent becomes slow or confused, review:

1. system prompt length;
2. number of assigned skills;
3. memory length;
4. enabled tool and MCP schemas;
5. chat history;
6. pipeline {prev} output or webhook payload size.

Reduce the source of unnecessary context before increasing the model window.