A progressive-loading memory architecture for a long-running personal AI agent, built on Claude Desktop + MCP.
Most personal-agent setups load their entire context file set on every session. That works until the knowledge base outgrows the budget — then every trivial exchange pays for the full corpus. OpenClaw separates context by access frequency instead of topic, and routes the rest in on demand.
Boot payload went from 37KB to roughly 6KB with no loss of recall.
- AGENTS.md — the architecture, tier model, and routing mechanism
- SOUL.md — persona and operating principles
- content/ — writing and content strategy
Personal data lives in a separate private repository. This one is config, architecture, and portfolio content only.
Claude Desktop · MCP (filesystem + shell) · Markdown as the only datastore · git as the only sync layer
Built by Shawn Tseng — .NET / Azure engineer, Melbourne.