Argus Agent is a memory-first AI system designed for long-context personal intelligence, offering robust recall, temporal and quantitative reasoning, and tool integration. It serves as an inspectable, open alternative to existing memory agents, prioritizing durable local storage, precise attribution, and self-correcting retrieval for enhanced AI agent capabilities.

How It Works

Argus employs a memory-centric architecture, moving beyond simple vector search to a full memory reasoning loop. It integrates local FAISS vector memory with distinct semantic, episodic, and procedural memory layers. A hybrid retrieval system combines vector search with keyword matching, enhanced by HyDE-style query expansion and dynamic recall depth. Crucially, it enforces strict temporal grounding and numeric attribution, featuring a self-correcting fallback retrieval mechanism and LangGraph orchestration to ensure accurate, contextually relevant responses and address common RAG failures.

Quick Start & Requirements

• Primary install / run command: Clone the repository, then execute scripts/setup_argus.py (using python3 on macOS/Linux or powershell on Windows). After setup, run the argus command from any directory.
• Non-default prerequisites and dependencies: Python 3.11+, an OpenAI API key. Optional: Telegram bot token, cloudflared (automatically installed by the setup script), OpenAI Codex CLI.
• Estimated setup time or resource footprint: The setup script automates environment creation, dependency installation, and configuration. Initial setup involves editing the .env file.
• Links: Repository: https://github.com/quarqlabs/argus

Highlighted Details

• Local-first Memory: Operates entirely locally without external database dependencies like Supabase pgvector.
• Multi-Layered Memory: Supports semantic (facts), episodic (events), and procedural (rules) memory types.
• Advanced Retrieval: Utilizes hybrid vector/keyword search, HyDE query expansion, and dynamic recall depth.
• Temporal & Quantitative Reasoning: Strict protocols for handling time and numerical data, distinguishing storage vs. event time and ensuring numeric attribution.
• Self-Correcting Retrieval: Implements fallback search mechanisms when initial retrieval yields insufficient evidence.
• Coding Agent Delegation: Integrates with providers like Codex for durable, streamed coding tasks, supporting various commands for task management and configuration.
• Control Console: A Codex-style CLI provides a rich interface for interacting with the agent, managing jobs, and viewing status.

Maintenance & Community

The project is described as an "active OSS release candidate" (v0.5.0). Specific details regarding maintainers, community channels (like Discord/Slack), or active development roadmaps beyond the release candidate status are not explicitly detailed in the provided README.

Licensing & Compatibility

• License type: Apache License 2.0.
• Compatibility notes: The Apache 2.0 license is permissive, generally allowing for commercial use and integration into closed-source projects.

Limitations & Caveats

Version 0.5.0 is an active release candidate, indicating ongoing development and potential for future changes. Structured artifact extractors are currently disabled to focus on tuning memory quality. Running comprehensive benchmarks like LongMemEval-S can be costly, estimated at ~$2,500 for a full run. The current focus is on single-user local memory, with multi-user serving planned for future development.