Rolling context, working across many sessions
How experienced AI operators build continuity even without a real memory system.
The problem
AI has no memory between sessions. Yesterday your project chat was full of context, today it's all gone. Level 4 solves that with memory systems. Until then, there's a simpler trick.
Rolling context
At the end of each session you ask the model: "Summarize what we decided today, where we stand, what comes next. 10 to 15 lines."
The model gives you a compact state-dump. Copy that into a notes file (or better: the tool of your choice).
On the next start you paste the state back as the first message. The model is instantly up to speed. No 20-minute "where did we leave off".
What goes in
- Project name + what's being built
- Key decisions we took (and why)
- Open questions or blockers
- Next concrete step
That's 150 to 300 words. Always fits in the context window.
What doesn't go in
- Full conversation transcripts (too long, redundant)
- Emotional reactions ("we were frustrated"), irrelevant for the next model
- Too many details ("on April 23 at 14:37 X said"), only relevant if they keep mattering
Practical trick
Build yourself a template file: PROJECT_STATE.md with three sections: Context, Decisions, Next Steps. After each session you update the file. Before each new session you paste it in.
That's manual memory. Two minutes at the end + 10 seconds at the start. Saves hours of re-explaining.
Upgrade path
Once you've done this for three weeks and the state file gets unwieldy (multiple projects, many decisions), it's time for a real memory system. See Level 4.