MCP Memory Service
Jul 2026 - Present
Shared long-term memory for AI coding agents, served over the Model Context Protocol. Agents stop re-reading the same files and stop retrying dead ends.
- Three memory tiers: working (current plan, so two agents don't edit one file), episodic (past attempts checked against real commits) and semantic (what the codebase does).
- Led a small team early on and scoped 50+ issues. Now I build it on my own. Built an A/B harness on SWE-bench Verified with Modal grading to measure cost per resolved issue with and without memory.
- Cut CI from 1h50m to 44m by running each repo, repeat and memory arm as its own parallel GitHub Actions job, then grading the patches on Modal.
- Python
- FastMCP
- SQLite
- React
- Modal
- Fly.io
- GitHub Actions