routine-memory

name	routine-memory
description	AgentDashboard Routine memory protocol. Use when a session is started by a scheduled, webhook, or plugin-triggered Routine and has a routine:// VFS mount; read trigger facts and durable Routine/entity memory for cross-trigger continuity, then update only concise future-use facts, decisions, open items, or entity outcomes.

Routine Memory

Use this skill when the current session was started by a Routine and a routine:// VFS mount is available.

Routine memory is a compact, durable working memory for the automation rule. It is not a transcript dump. Use it only when the current task benefits from cross-trigger continuity.

For the full file semantics, read references/memory-model.md.

Read Order

1. Read routine://memory/brief.md to understand the Routine's durable purpose.
2. Read routine://memory/open-items.md when the task may continue prior work.
3. If routine://current/execution.json has an entity_key, read routine://entities/{entity_key}/brief.md and routine://entities/{entity_key}/open-items.md.
4. Read routine://current/trigger.json for this trigger's payload and source facts.

Update Rules

• Update memory/facts.md only with durable facts that should survive future triggers.
• Update memory/decisions.md only with decisions that affect future behavior.
• Update memory/open-items.md with unresolved follow-up work.
• Update entities/{entity_key}/last-run.md after entity-scoped work when the outcome matters for the next trigger.
• Keep entries concise. Prefer summaries, facts, decisions, and recovery points over raw logs.

Boundaries

• Treat current/* as read-only trigger facts.
• Do not copy large payloads, transcripts, or tool logs into memory.
• Do not invent facts. Mark uncertain observations as pending confirmation.
• Use the Routine prompt as the task instruction and Routine memory as supporting context.