Memory System

Teleton uses a hybrid RAG architecture combining vector embeddings (sqlite-vec) with full-text search (FTS5) for intelligent context retrieval.

Hybrid Architecture

The memory system combines two complementary search methods:

Component	Technology	Purpose
Vector Search	`sqlite-vec`	Semantic similarity matching
Full-Text Search	`FTS5`	Exact keyword matching
Database	`better-sqlite3`	WAL mode for concurrent access

By default, Teleton uses local ONNX embeddings for privacy and offline operation:

config.yaml

memory:
  embeddings:
    provider: local          # local (ONNX) | voyage
    model: all-MiniLM-L6-v2  # Lightweight, fast model

For higher quality embeddings, you can use Voyage AI:

config.yaml

memory:
  embeddings:
    provider: voyage
    model: voyage-2
    apiKey: ${VOYAGE_API_KEY}

AI-powered memory compaction keeps the context relevant by summarizing and consolidating old entries:

config.yaml

memory:
  autoCompact: true        # Enable AI compaction
  compactThreshold: 1000   # Trigger after N entries

All memory data is stored in a single SQLite database:

Path

~/.teleton/memory.db