tdmcp

English

Português

English

Português

Appearance

Session profile & corpus learning ​

By default every conversation starts cold. This arc lets tdmcp remember your taste across sessions: a persistent profile the AI loads in one read, fed by tools that learn your conventions from a live project, from your own library of notes, or from a standing record of your preferences.

Reach for this when you find yourself repeating the same instructions — "always magenta on black", "name containers in snake_case", "I prefer the particle flock over the GPU field" — and want the assistant to start from your style instead of a generic default.

The session profile ​

load_session_profile (the AI-layer tool) reads or initialises ~/.tdmcp/session-profile.json — a cross-session snapshot that caches the latest outputs of the learning tools below. It returns a unified JSON with a loaded_at timestamp and style_memory, recent_work, conventions and corpus_style sections, creating sensible defaults if no file exists. Override the path with TDMCP_SESSION_PROFILE_PATH.

The same data is exposed read-only as the tdmcp://session/profile MCP resource, so an agent can pull all your preferences in a single resource read at the start of a turn.

"Load my session profile so you know my style before we build."

The learning tools ​

These four tools (the vault group — they need an Obsidian vault via TDMCP_VAULT_PATH) populate the profile. None of them mutate your TouchDesigner project:

  • style_memory reads, updates or summarises a standing Memory/style.md note — your long-lived record of palettes, default energy, banned moves, favourite generators and naming/layout conventions. Modes: show (a compact one-line context for the LLM), read (full structured), update (field-wise merge).
  • learn_conventions walks a live TouchDesigner subtree read-only and extracts your house conventions — naming case, colour tags, container shapes, parameter defaults, layout direction — writing them to Memory/conventions.md and optionally merging confident signals into Memory/style.md. No TD mutations.
  • learn_from_my_corpus is the offline companion: it walks your vault corpus (Recipes, Components, Looks, Setlists, Moodboards) and distils palette, naming, recipe-shape and parameter preferences into Memory/corpus_style.md. No TouchDesigner required — pure filesystem read.
  • recall_similar_work ranks your memory notes by similarity to a visual goal (query tokens plus tag/operator overlap) and returns the closest prior recipes, params and prompts so the agent can reuse them.

"Learn my conventions from /project1, then learn from my vault corpus, and update my style memory with anything you're confident about."

A typical flow ​

  1. Configure a vault (TDMCP_VAULT_PATH) and build a few things you like.
  2. Run learn_conventions on a project you're proud of and learn_from_my_corpus over your saved looks; both write to Memory/*.md.
  3. Run load_session_profile to fold those into ~/.tdmcp/session-profile.json.
  4. In future sessions, the AI reads tdmcp://session/profile (or you say "load my profile") and builds in your style from the first prompt.

Corpus learning above is about your own taste. tdmcp also ships two opt-in, local-only retrieval libraries that broaden the well the AI draws from — both gated behind TDMCP_RAG_ENABLED=1 and off by default:

  • Creative RAG — a curated repertoire of techniques and references, exposed as tdmcp://creative/cards/{id} and tdmcp://creative/search. See Creative RAG.
  • Project RAG — a local index of TouchDesigner projects/components (also needs TDMCP_PROJECT_RAG_ENABLED=1), exposed as tdmcp://project/cards/{id}, tdmcp://project/search and tdmcp://project/sources. See Project RAG.

Both are read-only and always carry source URL, license and rights notes on every result. They are documented in full on their own pages; this guide covers the always-on session-profile path.

See also ​