otto-agent
A transparent, local-first agent runtime with memory, skills, tasks, subagents, and context governance.
otto-agent is a personal AI assistant runtime I built around the idea that useful assistants need durable memory, structured tool use, and explicit operational constraints, not only model capability.
The assistant supports three entry points:
- CLI for fast local workflows
- Web dashboard for visibility and control
- Voice entry for lightweight interaction
Key system capabilities:
- Long-term memory and Knowledge RAG backed by
PostgreSQL,pgvector, full-text search, andHNSW. - Progressive Skill loading and
MCP-based tool extension for controlled capability growth. - Layered context governance with observation compression, history trimming, and summary-based context recovery.
- Permission-aware file and terminal operations with trace hooks, evaluation harnesses, and runtime telemetry export.
From an engineering standpoint, the project is where I concentrated work on agent harnesses, context engineering, memory design, and evaluation infrastructure. It is less about a single model and more about turning model calls into a system that can be inspected, constrained, and improved over time.
The stack includes Python 3.11+, Anthropic/OpenAI SDKs, PostgreSQL, pgvector, Pytest, and uv.