Crate loci 

Cognitive memory for AI agents — persistent, structured, cross-session memory via MCP.

Loci is an MCP server that gives AI agents a memory system inspired by cognitive science. Memories are stored in four types, each with different scoping, decay rates, and lifecycle behaviors:

Type	Purpose	Default Scope	Decay
Episodic	Events, decisions, session logs	Group	Fast (0.95/cycle)
Semantic	Facts, knowledge, preferences	Global	Slow (0.99/cycle)
Procedural	Workflows, patterns, how-to	Global	Slow (0.99/cycle)
Entity	People, places, projects	Global	Slow (0.99/cycle)

Architecture

• Storage: SQLite with FTS5 for keyword search and sqlite-vec for vector search
• Embeddings: Local ONNX Runtime with all-MiniLM-L6-v2 (384 dimensions)
• Search: Hybrid vector + BM25 keyword search merged via Reciprocal Rank Fusion
• Transport: MCP over stdio (primary) or Streamable HTTP/SSE

Modules

• config — Configuration loading from TOML files and environment variables
• db — SQLite database initialization, schema, migrations, and health checks
• embedding — Text-to-vector embedding pipeline via ONNX Runtime
• memory — Core memory engine: store, search, forget, relations, and maintenance

Modules

config

Configuration loading and management.

db

Database initialization and health checking.

embedding

Text-to-vector embedding pipeline.

memory

Core memory engine — storage, search, relations, and maintenance.

tools

MCP tool handlers — bridges JSON-RPC tool calls to the memory engine.