This project addresses the common issue of AI agents suffering from "amnesia," requiring users to repeatedly provide context. TrueMemory offers an automatic, 100% local, and persistent memory solution for AI agents, enhancing their contextual understanding and user experience without compromising data privacy. It is designed for AI developers, researchers, and power users integrating with tools like Claude Code, Gemini CLI, and Cursor.

How It Works

TrueMemory employs a retrieval-centered architecture, automatically capturing relevant conversation snippets and storing them in a single, local SQLite file. It functions as a "living memory," adapting over time by resolving contradictions and updating facts, thereby improving recall accuracy with continued use. This approach bypasses the need for manual tagging or complex prompt engineering, providing instant, contextually relevant information to AI agents without relying on cloud services or API keys (except for optional Pro tier features).

Quick Start & Requirements

Installation for AI tool integration is achieved via a bash script (curl -LsSf https://raw.githubusercontent.com/buildingjoshbetter/TrueMemory/main/install.sh | sh) or PowerShell equivalent, which installs dependencies in an isolated environment and downloads approximately 1.5GB of AI models. Setup takes 3-5 minutes. For developers, a Python library is available via pip install truememory. Base and Pro tiers require 4GB+ RAM. The Pro tier necessitates an LLM API key for its HyDE search functionality.

Highlighted Details

• Achieves high benchmark scores: 93.0% on LoCoMo and 92.0% on LongMemEval with the Pro tier.
• Offers state-of-the-art (SOTA) performance on BEAM-1M at 76.6%.
• Provides three tiers: Edge (lightweight, CPU-only), Base (fully offline, 4GB+ RAM), and Pro (enhanced search with LLM API key).
• All data is stored locally in ~/.truememory/memories.db, with Edge and Base tiers making zero external network calls.
• The architecture is backed by the peer-reviewed research paper "Storage Is Not Memory: A Retrieval-Centered Architecture for Agent Recall" (arXiv:2605.04897).

Maintenance & Community

The project lists several contributors, including buildingjoshbetter and SoilChang, and is associated with Sauron Labs. Community engagement is encouraged through X (formerly Twitter) accounts (@Building_Josh, @Sauron_Labs) and GitHub Discussions.

Licensing & Compatibility

TrueMemory is licensed under AGPL-3.0. This license permits free use for personal and research purposes but requires a separate commercial license for commercial applications, due to its copyleft provisions.

Limitations & Caveats

The advanced HyDE search feature in the Pro tier requires an external LLM API key. The AGPL-3.0 license imposes significant copyleft obligations, meaning any modifications or derivative works distributed must also be made available under the AGPL-3.0 license, potentially complicating integration into proprietary commercial software without a dedicated commercial license.