Drawer re-render · Memory surface
Memory — agent-managed campaign state
The Memory drawer slides out from the left rail of the Nagato widget. The agent updates campaign memory through a validated action boundary — never freeform writes. Objectives, TODOs, choices, and session recaps are all scoped per game. UI add/edit actions compose chat prompts; the agent performs the actual mutation.
▲ Try the filters — click type pills or search the memory.
How memory works
🛡️ Validated action boundary
The model can only touch campaign memory through a validated action allowlist, capped per response, scoped to the current game. No freeform writes, ever. The UI drafts prompts; the agent performs mutations.
🎮 Per-game isolation
Memory is keyed to the detected game. Cross-game access is a hard permission error, tested. Starting a different game starts a fresh campaign context.