Synaptic Memory
Synaptic Memory is a knowledge-graph and MCP tool server for LLM agents that builds a graph of documents, chunks, database rows and their relations instead of only chunk embeddings, so an agent can follow relations rather than rely on similarity search alone.
A single command builds a small SQLite-backed graph from a data source and runs example searches without calling any LLM at indexing time, auto-detecting file format or database schema, generating an ontology profile and building foreign-key edges. Presets keep common configuration compact: a local preset that is fully deterministic, a rag preset that reads embedding and reranking endpoints from configuration, and an agent preset that adds deterministic component bridging for multi-turn exploration.
The project's stated case against plain retrieval-augmented generation is that chunk similarity alone cannot answer a question that requires combining two separately stated facts, where a graph traversal can. Live database sync keeps the graph current through change-data-capture style watermarks or row hashing, and office file formats such as PDF, DOCX, PPTX, XLSX and HWP are supported through an optional add-on package. Korean full-text search is built in.
It is written in Python and installed from PyPI.