Every developer who uses AI assistants has experienced this: the model does something wrong, you correct it, and then… nothing. That correction evaporates. Next session, same mistake.
I’m building Sessions to fix this.
The Problem
AI assistants are stateless between sessions. They don’t remember that you prefer uv over poetry, or that your project uses docker-compose instead of running services directly. Every session starts from zero.
The Solution: In-Session Annotations
Sessions introduces simple commands that let you annotate corrections in real-time:
User: Restart the dev server
Model: killall dev; go run ./dev
User: <interrupts>
User: /sessions:project We have docker-compose for the dev server
Model: Noted
Model: docker-compose up -d --force-recreate
Or for global preferences:
User: Create a new python project
Model: poetry init
User: <interrupts>
User: /sessions:global Use uv for python projects
Forward-Backward
The key insight is that learning happens in both directions:
Forward: Your annotations immediately affect the current session. The model adapts on the spot.
Backward: Post-session analysis finds instances where the model performed incorrect actions, records them, and feeds that context into future sessions.
Your corrections compound. The AI assistant you use tomorrow is shaped by every correction you made today.
Why This Matters
This isn’t about making AI smarter in general. It’s about making your AI assistant smarter about your workflow. Personal context that persists and improves.
The best tool is one that learns how you work.
Comments
0Want to join the discussion?
Login to comment