An AI companion that proactively initiates conversations and performs background tasks, Akashic Agent moves beyond passive Q&A. It intelligently determines when and what to message based on subscribed information sources, making it suitable for users seeking a more engaged and autonomous AI assistant. The agent can also execute background tasks during idle periods, enhancing productivity.

How It Works

Akashic Agent operates through a multi-faceted system. In passive mode, it processes incoming messages through a memory retrieval and tool-calling pipeline across six distinct phases. For proactive engagement, it adapts polling frequency and analyzes alert, content, and context data streams, using an LLM to decide whether to push information. A sophisticated memory system consolidates conversations into structured facts and semantic embeddings for efficient retrieval. When idle, the agent executes user-defined background tasks (SKILL.md) through its "Drift" feature, preventing wasted cycles. The plugin system allows deep integration and customization via PhaseModules, EventBus decorators, or tool registration.

Quick Start & Requirements

• Primary install: Clone the repository, navigate to the directory, and set up a virtual environment using uv (uv venv && uv pip install -r requirements.txt). Ensure uv is installed (pip install uv).
• Prerequisites: Python 3.12. API keys for LLM providers (DeepSeek, Qwen/DashScope) and potentially embedding models are required. Telegram token is recommended for the communication channel.
• Setup: Run uv run python main.py setup for an interactive setup wizard or uv run python main.py init for non-interactive initialization. Configuration is managed via config.toml.
• Run: Execute uv run python main.py to start the agent.
• Docs: Links to specific guides for proactive messaging, drift tasks, memory, and plugins are available within the README.

Highlighted Details

• Proactive push frequency dynamically adjusts based on user interaction, from 8-minute intervals after a chat to 1-minute intervals during inactivity.
• The memory system consolidates conversations into structured facts (HISTORY.md, PENDING.md) and utilizes a vector database (memory2.db) for semantic search, with a separate MEMORY.md for prompt caching.
• The plugin system offers multiple intervention points, including PhaseModule chains, EventBus decorators, and tool interception/registration.
• The "Drift" feature allows the agent to autonomously execute background tasks defined in SKILL.md during idle periods.

Maintenance & Community

No specific details regarding maintainers, community channels (like Discord/Slack), or project health signals were found in the provided README.

Licensing & Compatibility

The README does not explicitly state the project's license. Compatibility notes suggest the project performs best with DeepSeek v4 Flash and Qwen models, with performance not guaranteed for other models, particularly those with poor XML output capabilities.

Limitations & Caveats

The project's performance is optimized for and not guaranteed with LLMs other than DeepSeek v4 Flash and Qwen, especially those struggling with structured output formats like XML. Direct image processing is not supported by the primary DeepSeek model, relying instead on Vision-Language (VL) tools.