The problem_
Most personal assistants forget you the moment the session ends. I wanted one that learns my patterns, remembers context across weeks, and actually executes tasks — not just suggests them.
What I built_
- Multi-agent assistant for task planning and workflow automation.
- Contextual decision-making using short-term and long-term memory.
- Features — reminders, information lookup, and task execution.
- Extensible architecture — drop-in agents for new capabilities with scalable autonomous decision support.
Stack_
- LangGraph for orchestration
- Qwen3 as the reasoning model
- Python backend, lightweight React front-end
- Vector store for long-term memory
Lessons_
Long-term memory is a feature, not a backend. Exposing what the assistant remembers — and letting the user edit it — turned out to be the single biggest trust-builder. Guardrails around task-execution agents were essential.
─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─ ─