@mise-en-scene/taste-memory

Aesthetic intelligence MCP server. Stores taste observations, design decisions, anti-patterns, and reference analyses in SQLite. Builds taste profiles that persist across sessions — the agent develops real preferences over time.

MCP Tools

| Tool | Input | Description | |------|-------|-------------| | taste_remember | type, scope, data, reasoning?, project? | Store a taste observation | | taste_recall | query, type?, scope?, project?, limit? | Search taste memory by text | | taste_profile | scope (user / project / merged), project? | Get aggregated taste profile | | taste_decide | chosen, rejected, reason, project? | Record a design decision | | taste_reference | project, analysis, image_path? | Store a reference analysis | | taste_system | project, tokens? | Get or set design system tokens |

Usage

As an MCP Server

The package runs as a stdio MCP server. Configure it in your .mcp.json:

{
  "mcpServers": {
    "mise-en-scene-taste-memory": {
      "command": "npx",
      "args": ["@mise-en-scene/taste-memory"]
    }
  }
}

Or run directly:

npx mise-en-scene-taste-memory

Environment Variables

| Variable | Default | Description | |----------|---------|-------------| | MISE_TASTE_DB | ~/.mise-en-scene/taste.db | SQLite database path |

Programmatic Usage

import { createDatabase, closeDatabase } from '@mise-en-scene/taste-memory/db';
import { TasteStore } from '@mise-en-scene/taste-memory/store';
import { TasteSearch } from '@mise-en-scene/taste-memory/search';
import { buildProfile } from '@mise-en-scene/taste-memory/profile';

// Initialize
const db = createDatabase('/path/to/taste.db');
const store = new TasteStore(db);
const search = new TasteSearch(db);

// Store observations
store.remember('palette', 'user', {
  dominant: '#1a1a2e',
  accent: '#e94560',
}, { reasoning: 'Dark moody palette with warm accent' });

store.remember('anti-pattern', 'user', {
  rejected: 'Roboto',
  context: 'too generic',
}, { reasoning: 'AI default font' });

// Record decisions
store.decide(
  { layout: 'asymmetric' },
  { layout: 'centered' },
  'Asymmetric creates visual tension',
  'wine-label'
);

// Store reference analysis
store.storeReference('wine-label', {
  colors: ['#f5f0e8', '#c9a96e'],
  typography: 'transitional serif + humanist sans',
  mood: ['luxury', 'warm'],
  cultural_context: 'French wine label tradition',
});

// Search
const results = search.textSearch('editorial dark premium');
const decisions = search.searchDecisions('layout', 'wine-label');

// Build profile
const profile = buildProfile(store, 'merged', 'wine-label');

// Cleanup
closeDatabase();

Data Model

Taste Types

| Type | What It Captures | |------|-----------------| | palette | Color choices — dominant, secondary, accent | | typography | Font preferences, pairings, classifications | | layout | Composition patterns, grid preferences | | texture | Surface treatments — grain, gradients, shadows | | mood | Emotional descriptors — "editorial", "warm", "refined" | | anti-pattern | Things to avoid — "generic gradients", "Roboto" |

Scopes

| Scope | Persists Across | |-------|----------------| | user | All projects — your personal aesthetic | | project | Single project — project-specific choices | | merged | Both combined, project overrides user |

Database Schema

Four tables in SQLite with WAL mode:

• taste_entries — Observations with type, scope, JSON data, reasoning, timestamps
• taste_decisions — Chosen vs rejected options with reasoning
• taste_references — Reference image analyses with structured extraction
• design_systems — Design tokens per project (upsert on conflict)

Search

Text search uses token-based similarity scoring (Jaccard-style overlap). Queries are tokenized, matched against stored entry data and reasoning fields, and ranked by relevance.

// Search with filters
const results = search.textSearch('warm editorial', {
  type: 'palette',
  scope: 'user',
  limit: 10,
});

Development

npm test                # Run tests (24 unit + 2 integration)
npm run build           # Compile TypeScript
npm run test:watch      # Watch mode