AI agents
Jimmy agent platform
A TypeScript multi-agent system on the Claude Agent SDK, with long-term memory and scheduled jobs.
- Status
- Deployed
- Role
- Design and build
- When
- 2026
- Working memorySession state for the task at hand
- Episodic memoryLightRAG over PostgreSQL with pgvector
- Semantic memoryCurated notes, consolidated nightly
Three memory layers behind each agent.
Most chat assistants start from zero every session. The goal was agents that keep memory between sessions and run scheduled work on their own.
What I built
- Seven agent personas with their own prompts and workspaces, led by an orchestrator called Jimmy.
- Memory in three layers: session state, LightRAG over PostgreSQL with pgvector, and curated notes consolidated nightly.
- A heartbeat service running 10+ scheduled jobs, and a Hono web dashboard with streaming chat.
- Model routing with fallbacks across providers, and Playwright MCP for browser tasks.