@steno-ai/engine

The core memory engine for Steno. Handles extraction, retrieval, knowledge graphs, salience scoring, contradiction detection, and more.

Install

npm install @steno-ai/engine

Architecture

Conversation / Raw Text
        |
        v
  ┌─────────────────────────────────────────┐
  │         Extraction Pipeline              │
  │  heuristic -> LLM -> dedup -> contradict │
  │  -> entity extraction -> embed -> store  │
  └─────────────────────────────────────────┘
        |
        v
  ┌─────────────────────────────────────────┐
  │         Storage (Facts + Graph)          │
  │  facts, entities, edges, embeddings      │
  └─────────────────────────────────────────┘
        |
        v
  ┌─────────────────────────────────────────┐
  │         6-Signal Retrieval               │
  │  vector + keyword + graph + temporal     │
  │  + salience decay + trigger matching     │
  │  -> fusion ranking -> reranking          │
  └─────────────────────────────────────────┘

Usage

Extraction Pipeline

import { runExtractionPipeline } from '@steno-ai/engine';

const result = await runExtractionPipeline(
  {
    storage,       // StorageAdapter (sqlite, supabase)
    embedding,     // EmbeddingAdapter (openai, openai-compat)
    cheapLLM,      // LLMAdapter for fast extraction
    smartLLM,      // LLMAdapter for complex extraction (optional)
    embeddingModel: 'text-embedding-3-small',
    embeddingDim: 1536,
  },
  {
    tenantId: 'tenant_1',
    scope: 'user',
    scopeId: 'user_123',
    sourceType: 'conversation',
    data: [
      { role: 'user', content: 'I just moved to San Francisco' },
      { role: 'assistant', content: 'Welcome to SF!' },
    ],
  },
);

console.log(result.facts);    // extracted memory facts
console.log(result.entities); // knowledge graph entities
console.log(result.edges);    // entity relationships

Search (6-Signal Retrieval)

import { search } from '@steno-ai/engine';

const results = await search(
  {
    storage,
    embedding,
    cache,                     // optional CacheAdapter
    rerank: true,              // embedding-based reranking
    salienceHalfLifeDays: 30,  // decay curve
  },
  {
    tenantId: 'tenant_1',
    scope: 'user',
    scopeId: 'user_123',
    query: 'where does the user live?',
    limit: 10,
    includeGraph: true,
  },
);

for (const r of results.results) {
  console.log(r.content, r.score);
}

Adapter Interfaces

The engine defines these interfaces -- bring your own implementations or use the official adapters:

| Interface | Official Adapters | |-----------|-------------------| | StorageAdapter | @steno-ai/sqlite-adapter, @steno-ai/supabase-adapter | | LLMAdapter | @steno-ai/openai-adapter, @steno-ai/openai-compat-adapter | | EmbeddingAdapter | @steno-ai/openai-adapter, @steno-ai/openai-compat-adapter | | CacheAdapter | @steno-ai/cache-adapter |

Key Modules

• extraction/ -- Heuristic + LLM fact extraction, deduplication, contradiction detection, entity extraction
• retrieval/ -- Vector search, keyword search, graph traversal, salience scoring, fusion ranking, reranking
• salience/ -- Time-decay scoring, feedback boosting
• sessions/ -- Session lifecycle management
• feedback/ -- Feedback tracking and fact score adjustment
• profiles/ -- User profile aggregation
• scratchpad/ -- Working memory / scratchpad updates

Part of Steno

The memory layer for AI agents.