Forgetful Search Functionality

December 3, 2025 ยท View on GitHub

This section details Forgetful's search functionality and associated configuration to allow you to tailor the service to your use case.

Auto-Linking Pipeline

When a new memory is created, Forgetful automatically builds knowledge graph connections:

Memory Auto-Linking Flow

How it works:

  1. Encode - The memory's title, content, context, keywords, and tags are combined and converted to a vector embedding
  2. Store - The embedding is stored in the Forgetful database alongside the memory
  3. Similarity Search - Existing memories are searched for semantic similarity
  4. Cross-Encoder Reranking - Candidate matches are scored by the cross-encoder for precision
  5. Auto-Link - Memories exceeding the 0.7 similarity threshold are automatically linked (bidirectional)

This creates a self-organizing knowledge graph where related concepts connect without manual intervention.

Embedding Providers

FastEmbed

For local embeddings and re-ranking we support the use of the excellent embedding solution developed by Qdrant.

Google

We also support Google Embedding models available via the Gemini API

Azure

Support for the Azure Foundary OpenAI embeddings is now added as well.

Configuration

The following configuration options are available for search