Agents, Skills, and Memory

This page helps you make agents reliable: clear prompts, scoped tools, reusable skills, and memory that stays useful over time without exposing every page to every agent.

Agent profile basics

  • One profile can serve multiple channels when behavior should stay consistent.
  • Use separate profiles for distinct workflows (support triage, release assistant, repository analyst).
  • Select model/provider per profile based on task requirements and plan limits.
  • Assign only the channels, sources, and memory access the profile actually needs.

System prompt structure that works

# Recommended prompt structure

Role: You are a support operations assistant.

Scope: Use linked workspace sources only; do not invent missing data.

Style: Be concise, list actions first, then rationale.

Safety: Confirm destructive actions before execution.

Citations: Reference source/tool results when factual claims are made.

Keep prompts explicit and operational. Avoid long prose and repeated rules.

Skills

  • Use skills for reusable instruction blocks (triage format, incident response template, release checklist).
  • Set priorities so high-value instructions always win when conflicts exist.
  • Use global skills as shared templates and workspace skills for team-specific guidance.
  • Treat skills as versioned operational playbooks and review periodically.

Memory pages

  • Create pages for durable facts: architecture notes, escalation policies, onboarding runbooks.
  • Avoid storing volatile one-off chat artifacts as long-term memory.
  • Prefer short, structured markdown sections over long unstructured text.
  • Use tags so agents and humans can keep memory organized by domain or workflow.
  • Review revision history and clean stale content regularly.

Workspace memory is shared workspace knowledge. Personal memory is separate: it belongs to the signed-in user and is managed from My Memory in account settings. Use personal memory for user-specific preferences, conventions, projects, decisions, and notes that should travel with that user rather than becoming workspace-wide guidance.

Suggested memory page template

Title: Escalation policy

Owner: Platform team

Last reviewed: 2026-03-01

Rules:

- Sev1: page on-call immediately

- Sev2: create incident thread + assign owner

- Sev3: next business day follow-up

Memory access modes

  • No memory access: the agent cannot read workspace memory pages.
  • All workspace memory: the agent can use any memory page in the workspace.
  • Only pages matching allowed tags: the agent can only use pages that share at least one allowed tag.

Tag-scoped access is the safest option when one workspace contains multiple teams or workflows. It lets you keep one shared memory library while still limiting each agent to the pages that match its allowed tags.

How tagged memory restrictions work

  • Tag pages consistently, for example support, billing, release, or incident.
  • When an agent is tag-scoped, it only reads pages that overlap with its allowed tags.
  • When a tag-scoped agent creates or updates memory, the page must keep at least one allowed tag.
  • If an agent should not use workspace memory at all, set its memory access to No memory access.

Operational tips

  • Bind only required sources/tools to reduce noise and context load.
  • Start with one strong profile before creating many specialized variants.
  • Use scheduled tasks to keep memory current (for example weekly summary refresh).
  • Decide on a tag taxonomy early so memory stays easy to scope as the workspace grows.
  • Audit agent output quality after changing prompt, model, or skill priority.