Memoria 1.0.0 Released: Local, Model-Agnostic Memory System for LLMs

Memoria 1.0.0 has been released as a local-first, LLM-agnostic memory system designed for local LLM applications without dependence on cloud services or specific models. The system operates independently of an LLM and is installable via pip install kitzkatz-memoria.

Performance and Hardware Efficiency

The developer benchmarked Memoria on a low-spec machine (Intel Celeron N4020, 3.7 GiB RAM, no GPU) to demonstrate its efficiency. On the LongMemEval-S benchmark, the system returned results for 468 out of 470 retrieval-evaluable questions.

Key metrics include:

• Recall@1: 89.8%
• Recall@5: 97.9%
• Recall@10: 98.9%
• Recall@50: 99.6%
• Session NDCG@10: 0.9257

Memory usage remained low, with peak RSS for the full workload at 2.65 GiB and average peak RSS per query around 580 MiB.

Retrieval Architecture

Memoria employs a multi-signal retrieval approach rather than relying solely on vector databases. It runs FAISS, BM25, graph retrieval, phrase matching, attribute retrieval, and temporal retrieval in parallel. These signals are fused and processed through multi-signal ranking. Temporal retrieval is independently implemented to allow for separate measurement and tuning.

Features and Integration

The release includes comprehensive integration options and developer tools:

• Ingestion: Support for GitHub repositories and Obsidian vaults.
• Interfaces: CLI, TUI, GUI, and API.
• Protocols: MCP (Model Context Protocol) support.
• Extensibility: A plugin system with 11 subsystems and 34 hooks, including an interactive plugin generator.
• Storage: Persistent local storage.