Startup QuarqLabs has open-sourced version v0.4.0 of its core project, the Quarq Agent. The project aims to give AI agents a long-term memory architecture for "continuous learning," addressing common pain points like forgetfulness, poor long-span reasoning, and hallucination. At the same time, QuarqLabs announced it is shifting its strategic focus entirely to robotics infrastructure. The project will be archived as an open-source research asset, and the company will no longer actively maintain it.
https://twitter.com/quarqlabs/status/2059320863070286177
The Quarq Agent introduces a novel four-layer modular long-term memory system: query, storage, reasoning, and learning. In the query layer, the system expands questions into multi-perspective retrieval hypotheses in the background, combining local FAISS vector search with keyword search. The storage layer categorizes memories into semantic (preferences and facts), episodic (event history), and procedural (behavioral instructions and formatting rules). The first two are backed by a local vector database and JSON, while the third is maintained by separate rule sets. In the reasoning layer, explicit guardrails prevent hallucinations by strictly distinguishing event time from storage time, isolating entity associations, and proactively admitting when information is insufficient. The learning layer runs an independent model asynchronously in the background to add, delete, update, and deduplicate memories without adding interaction latency.
On the ecosystem and deployment front, the Quarq Agent focuses on minimal local deployment and comes with out-of-the-box skills like Gmail, Google Calendar, and structured PDF report generation. Developers can extend it dynamically by adding Python scripts in the tools directory. As a fully open-source, thoroughly designed memory-first agent reference implementation, it provides a high-quality engineering foundation for building long-term companion agents.