Notes — Agent Plot Thread Tracker

Notes is the orchestrator's per-chat store for what the agent (as plot author) wants to make happen later. Foreshadowing planted, promises made by characters, chapter outlines written ahead, mystery-genre solutions drafted before the investigation — anything that captures an obligation the agent has committed to honoring in a future turn.

It is deliberately distinct from memory-graph:

Subsystem	Answers	Author
memory-graph	"What has already happened" — objective record of scenes, character states, locations	Auto-extracted; user can curate
notes	"What does the agent (as author) want to happen next" — planted but not yet triggered	Agent itself, via tools during the loop

You (the player) can view, close, edit, and hard-delete notes through the Notes panel — but you are not the primary author. If you want to inject plot direction, use lorebook entries or memory-graph always-injected nodes; notes is for the agent's own bookkeeping.

Tools the agent uses

Tool	Behavior
note_open(text)	Open a new note. Returns its id. The note shows up in the agent's "## Open Notes" system block every subsequent run, until closed.
note_close(id, reason?)	Close a note by id with an optional one-line reason. The note disappears from the prompt block; the closure stays in the archive (visible in the panel).

Open notes are always visible in the agent's system prompt — there is no auto-LRU, no token budget, no auto-expiry. The agent is responsible for its own note hygiene: if it opens 80 stale notes and never closes them, it pays the prompt-budget cost itself. The orchestrator does not bail it out.

Anti-pollution principle

Notes is for genuine commitments, not a turn diary. A polluted notes list (entries that were never real obligations, or stale entries that should have been closed) drags every subsequent round. The director mode default sub-agents enforce this:

• notes_pickup_scout runs pre-draft and picks only the ripe open notes for this turn's beat
• notes_curator runs post-draft, defaults to do nothing, closes only on explicit textual evidence in the draft, and opens new notes only when the draft itself committed to a future-payoff obligation

If you write your own director sub-agents or operate in loop mode without a curator, the same discipline applies — close more aggressively than you open.

The Notes panel

The Notes panel lives in the orchestrator extension's panel, scoped to the current chat. It shows two tabs:

• Open notes — currently active threads. Each row has Close / Edit / Delete actions.
• Closed notes — the archive. Closure reasons are shown beneath each row.

Per-row actions:

Action	Effect
Close this note	Same as the agent's note_close — a closure-reason prompt appears; submitting moves the note to Closed.
Edit note	Inline text edit. Status and id are preserved. Useful for typo / clarity fixes.
Delete note (permanent)	Hard delete from storage. Two-step confirm. Not the same as Close — Delete erases the entry; Close keeps it for the record.

The panel updates live as agents open / close notes via tools, and your panel actions update agents' next-turn prompts the same way.

Common questions

Q: Should I manually open notes in the panel to give the agent plot hints? No. Notes is the agent's instrument. If you want to seed plot direction, use a lorebook entry or a memory-graph always-injected node — those are designed as user-side authoring surfaces.

Q: Are notes shared across chats? No — each chat has its own notes store. Switching chats shows that chat's own open / closed lists.