Nagato Knowledge, Recall & Trace Plan

Multi-page implementation blueprint

Evolve Nagato from a response surface into a transparent, game-scoped companion: stable visual Recall, inspectable model/tool traces, compact agent-owned drawers, and a provenance-preserving local game knowledge library.

8
linked plan pages

Each workstream has schemas, UX, test seams and acceptance criteria.

22
LM Studio models detected

21 generative + 1 Nomic embedding model at the live local endpoint.

4
implementation waves

Start with usability/provenance, then ingestion and retrieval.

Product principles

Raw first

Persist immutable raw provider and tool outputs. Normalized cards and local-model explanations are derived views with explicit provenance.

Agent-owned mutation

Drawer + buttons compose editable prompts into Chat; validated actions perform mutations. Filters never mutate records.

Game-scoped privacy

Every screenshot, trace, document, chunk and embedding belongs to normalized game_id. Local files require explicit roots and previews.

Honest reasoning UX

Show reasoning only when the provider emits a supported reasoning channel. Never label status prose or reconstructed summaries as raw chain-of-thought.

Recommended order

1 Stable Recall + raw VLM2 durable run traces3 compact drawers4 source ledger5 install/manual ingestion6 hybrid retrieval

Key architectural decision

Do not bolt RAG, raw tool output, and Recall metadata into the JSON GameStore. Keep campaign records there for now, but add durable SQLite stores for traces and source artifacts, then expose narrow bridge commands to the UWP. The widget remains a renderer and intent composer.

Pages

Recall evidence

Selection stability, raw VLM, OCR, review and carousel provenance.

Reasoning & tools

Live fading tail, expansion, historical trace inspection and raw output.

Drawer UX

Filter + plus compression and icon swap.

Source ledger

Capture every GTC/tool source with versions and provenance.

Game ingestion

Walkthroughs, wikis, PDFs and install Extras onboarding.

Retrieval/RAG

FTS5 + Nomic embeddings + optional reranking.

Visual concepts

Faithful current/proposed shell renders, trace variants and interactive Recall tabs.

Live evaluation

Torment benchmark from 33 actual manual/chat turns, plus future game cohorts.

Implementation roadmap

TDD slices, file map, dependencies and release gates.