A game companion that remembers the campaign, not just the question.
Nagato lives inside the Windows Game Bar overlay. It detects which game is running, answers with spoiler-safe hints, screenshots and reads gameplay with a local vision model, keeps per-game memory of objectives, choices and progress — and backs every research answer with verified, tamper-evident evidence.
What it does today
Everything below is installed and running. Each page of this showcase re-renders the real UI — click around.
💬 Spoiler-safe conversations
Ask anything about the game in front of the overlay. A spoiler dial — none / hint / full — survives every code path, so "give me a nudge" never becomes a walkthrough.
📝 Notes drawer
Game-scoped saved notes, filterable, with agent-drafted prompts. The UI never mutates data directly — add/edit composes a chat prompt for the agent.
🔍 Research drawer
Cached source-backed research packs. Every successful web source becomes immutable, hash-verified, versioned evidence — tombstoned on delete, never silently resurrected.
🗂️ Memory drawer
Campaign records with filters: objectives, TODOs, choices, and session recaps. The agent updates memory through a validated action boundary — never freeform writes.
⚙️ Settings
Accent color, spoiler default, model routing, hotkeys, and Recall configuration — all live, all local. Source trust proposals require explicit approval.
📸 Recall
Automatic gameplay screenshots, captioned and OCR'd by a local vision model. Searchable timeline — with real screenshots from a Torment: Tides of Numenera playthrough.
🧾 Trace inspector
Every tool call streams into a fading trace tail. Any result opens a raw artifact inspector with local model explanations. Glass-box AI.
⚗️ Hybrid retrieval + benchmark gate
Chunked lexical + vector retrieval with spoiler budgets — and a benchmark gate that keeps it disabled until it beats the baseline without leaking spoilers. The system tells on itself.
📖 Game codex
Every character, place, item and choice encountered becomes a browsable codex with evidence trails — spoiler-safe by construction, because it only knows what has been seen.
The design rules that make it trustworthy
This isn't a chatbot bolted onto a game. A few invariants are enforced in code, checked by tests, and audited on every merge:
Raw evidence is authoritative
AI summaries are labeled, derived and disposable. The bytes the tools actually returned — hash-verified — are always one click away.
Everything is scoped per game
Conversations, memory, screenshots, evidence: all keyed to the detected game. Cross-game access is a hard permission error, tested.
Deletion means deletion
Deleted evidence is tombstoned, unsearchable, and cannot silently return — even if the same bytes are re-downloaded later.
Never plays the game for you
No input automation, no botting, no save editing. Nagato watches, remembers and advises. That's a hard product line.
Built by AI, directed by one human
This showcase and the product it describes are 100% AI-generated under human direction. V1 was built by Claude Code (Claude-Fable-5). V2 is built by Hermes (GLM-5.2) on Featherless. The product code itself is built and cross-reviewed by three AI harnesses.