DeepaData EDM SDK

The significance layer for AI memory.

EDM artifacts encode what mattered at capture time — emotional weight, recall triggers, identity thread, arc type — so every memory architecture gets a richer signal to retrieve against.

Why EDM

Most memory systems score relevance at retrieval time. EDM encodes significance at capture time.

The difference: a raw conversation chunk tells your retrieval system what was said. An EDM artifact tells it what mattered — and why it might matter again.

What this enables:

• Retrieve by emotional significance, not just semantic similarity
• Surface moments that keyword and vector search miss entirely
• Give memory platforms five retrieval axes instead of one
• Portable, governed artifacts that travel with the person across platforms

In a controlled retrieval comparison, EDM field routing outperformed raw vector similarity by 55.6 percentage points on significance-typed queries — the class of queries that have zero lexical overlap between the question and the answer. These queries require significance structure. Raw text cannot answer them.

The spec is MIT licensed and published on Zenodo. The SDK is the extraction layer.

Installation

npm install deepadata-edm-sdk

Quick Start

import { extractFromContent, createStatelessArtifact, validateEDM } from 'deepadata-edm-sdk';

// Set your Anthropic API key
process.env.ANTHROPIC_API_KEY = 'your-api-key';

// Extract EDM artifact from content
const artifact = await extractFromContent({
  content: {
    text: "Looking at this old photograph brings back so many memories of summer afternoons at grandmother's house...",
  },
  metadata: {
    subjectId: "vp-01HZ3GKWP7XTJY9QN4RD",
    jurisdiction: "GDPR",
    consentBasis: "consent",
  },
});

// Validate the artifact
const validation = validateEDM(artifact);
if (!validation.valid) {
  console.error('Validation errors:', validation.errors);
}

// For session use: create privacy-preserving stateless version
const stateless = createStatelessArtifact(artifact);

// The artifact is now ready for sealing with ddna-tools
console.log(JSON.stringify(artifact, null, 2));

Features

• LLM-Assisted Extraction: Uses Claude, OpenAI, or Kimi to extract emotional data from text and images
• EDM v0.7.0 Compliant: Full support for all 10 EDM domains
• Profile Support: Essential (~20 fields), Extended (~45 fields), or Full (all fields)
• Text + Image Fusion: Combines text narrative with image context
• Stateless Mode: Privacy-preserving mode for session use
• Schema Validation: Comprehensive Zod-based validation
• Crosswalk Mapping: Automatic mapping to Plutchik and HMD taxonomies

Profiles

EDM v0.7.0 introduces profile-aware extraction. Choose the profile that matches your use case:

| Profile | Fields | Use Case | |---------|--------|----------| | essential | ~20 | Memory platforms, agent frameworks, AI assistants | | extended | ~45 | Journaling apps, companion AI, workplace wellness | | full | 96 | Therapy tools, clinical applications, research |

// Essential profile - minimal extraction
const artifact = await extractFromContent({
  content: { text: "..." },
  metadata: { consentBasis: "consent" },
  profile: "essential",
});

// Extended profile - balanced extraction
const artifact = await extractFromContent({
  content: { text: "..." },
  metadata: { consentBasis: "consent" },
  profile: "extended",
});

// Full profile - all fields (default)
const artifact = await extractFromContent({
  content: { text: "..." },
  metadata: { consentBasis: "consent" },
  profile: "full",
});

See examples/ for complete profile artifacts.

API Reference

Extraction

extractFromContent(options)

Extract a complete EDM artifact from content.

const artifact = await extractFromContent({
  content: {
    text: "User's narrative...",
    image: "base64-encoded-image", // optional
    imageMediaType: "image/jpeg",  // optional
  },
  metadata: {
    subjectId: "vp-01HZ3GKWP7XTJY9QN4RD",      // optional, null in stateless
    jurisdiction: "GDPR",          // optional
    consentBasis: "consent",       // required
    locale: "en-us",               // optional
    visibility: "private",         // optional
    piiTier: "moderate",           // optional
    tags: ["family", "memory"],    // optional
  },
  model: "claude-sonnet-4-20250514", // optional
  provider: "anthropic",             // optional: "anthropic" | "openai" | "kimi"
  profile: "full",                   // optional: "essential" | "extended" | "full"
});

extractFromContentWithClient(client, options)

Same as above but with a provided Anthropic client.

Stateless Mode

createStatelessArtifact(artifact)

Create a privacy-preserving version for session use.

const stateless = createStatelessArtifact(artifact);
// - owner_user_id: null
// - associated_people: []
// - location_context: null
// - recall_triggers: []
// - retrieval_keys: []
// - exportability: "restricted"

isStateless(artifact)

Check if an artifact is in stateless mode.

validateStateless(artifact)

Validate stateless compliance with detailed violations.

Validation

validateEDM(artifact)

Validate against the complete EDM v0.7.0 schema.

const result = validateEDM(artifact);
// { valid: boolean, errors: ValidationError[] }

validateDomain(domain, data)

Validate a specific domain.

const result = validateDomain("constellation", artifact.constellation);

validateCompleteness(artifact)

Check field population and completeness.

const result = validateCompleteness(artifact);
// { complete: boolean, missingFields: string[], populationRate: number }

Helpers

createEmptyArtifact()

Create an empty artifact structure for manual population.

createAnthropicClient(apiKey?)

Create an Anthropic client (uses ANTHROPIC_API_KEY env var if not provided).

EDM v0.7.0 Domains

The SDK supports all 10 EDM domains:

| Domain | Description | Extraction | |--------|-------------|------------| | meta | Identity, provenance, consent | From metadata | | core | Anchor, spark, wound, fuel, bridge, echo, narrative | LLM | | constellation | Emotional classification, archetypes | LLM | | milky_way | Contextual grounding (people, place, event) | LLM | | gravity | Salience, weight, retrieval keys | LLM | | impulse | Motivational state, drive, coping | LLM | | governance | Jurisdiction, retention, rights | From metadata | | telemetry | Extraction confidence, model ID | Auto-populated | | system | Embeddings, indices (downstream) | Empty | | crosswalks | External taxonomy mappings | Auto-mapped |

Integration with ddna-tools

After extraction, seal the artifact for portability:

import { seal } from 'deepadata-ddna-tools';

const envelope = await seal(artifact, privateKey, did);
// Result: signed .ddna envelope

Interpretation Boundary

This SDK performs extraction, not inference. The LLM extracts what is explicitly stated or clearly implied in the source content. It does not:

• Infer psychological states beyond stated emotions
• Diagnose conditions or suggest treatments
• Make predictions about future behavior
• Generate content not grounded in the input

See docs/EXTRACTION_VS_INFERENCE.md for details.

EU AI Act Compliance

This SDK is designed for compliance with the EU AI Act. Emotional data extraction with interpretation constraints is considered lower risk than emotion inference systems. See docs/EU_AI_ACT_COMPLIANCE.md.

Schema Version

This SDK implements EDM v0.7.0. Key changes from v0.6:

• Added meta.profile field for conformance level declaration
• Profile-aware extraction with tailored system prompts
• Profile-aware confidence scoring
• Three conformance profiles: essential (24 fields), extended (50 fields), full (96 fields)
• Added Kimi K2 extractor support via MoonshotAI API

Works With Your Memory Stack

EDM artifacts are format-agnostic episodic memory records. The structured fields integrate naturally with the retrieval infrastructure you already use:

• Embedding-based retrieval — embed core.narrative for semantic similarity search against any vector store
• Temporal knowledge graphs — map associated_people and tether_type as graph nodes and edges; temporal_context and recurrence_pattern as edge properties
• Lexical search (BM25) — recall_triggers and retrieval_keys are designed for activation-path matching against conversation fragments
• Hybrid retrieval — combine any of the above; EDM fields give each axis a significance-weighted signal rather than raw text
• Significance activation — the DeepaData activation API returns ranked artifact IDs for the current conversation moment (deepadata.com/docs/activation)

EDM encodes the episodic layer. The retrieval architecture is yours.

See deepadata.com/docs for integration guides.

License

MIT