Memoria¶

Local-first, LLM-agnostic memory substrate with parallel hybrid retrieval, multi-signal ranking, declarative type routing, and a plugin-based architecture.

4GB RAM · CPU-only · No cloud · No API keys required · MIT licensed
Built solo. Designed to be extended rather than replaced.

Memoria is a configurable memory/retrieval substrate, not a chatbot-specific memory implementation. It stores memories, routes queries by memory type, retrieves candidates using parallel workers, ranks and finalizes results, constructs context, and exposes the system through CLI, TUI, GUI, and API interfaces.

Where To Start¶

I want to…	Read this
Install Memoria and run my first query	Getting Started
Understand every setting	Configuration
Build an adapter for a new dataset	Adapters
Understand how retrieval works	Retrieval
See benchmark methodology and results	Benchmarks
Use the HTTP API	API
Understand the architecture	Architecture
Write a plugin	Plugins

What It Is¶

Memoria is a fully local memory system for LLMs and other applications that need persistent contextual retrieval.

Memoria stores raw memories. Extraction produces indexes — entities, relationships, embeddings, importance — that annotate each raw memory. Retrieval returns the original memory, not a decomposed fragment of it.

Retrieval workers can include:

FAISS — semantic retrieval
BM25 — lexical retrieval
Graph — entity/relationship traversal
Phrase — phrase matching
Attribute — structured attribute retrieval
Fusion — combined retrieval strategies

Retrieval workers are coordinated through a blackboard/scheduler architecture with declarative completion policies rather than requiring the query handler to synchronously wait for every worker.

You own your data. No cloud. No subscription.

Design Goals¶

Local-first
LLM-agnostic at the memory layer
Retrieval responsible for finding candidates; ranking responsible for deciding usefulness
Independently replaceable retrieval workers
Extensible through plugins, not forks

Project History & Design¶

License¶

MIT