Every time my AI agent started a new session, it woke up blank.
No memory of yesterday's work. No context about ongoing projects. No awareness of decisions we'd made together.
I'd spend the first ten minutes of every session catching it up.
Explaining what we were building, what had already been tried, what didn't work.
It was exhausting.
The agent was capable, but it had no continuity. Every conversation was like meeting for the first time.
So I built a memory system. Not a database. Not a vector store. Just simple markdown files.
MEMORY.md for long-term curated memories. Things that matter over weeks and months. Decisions, lessons learned, important context.
Then daily files in a memory folder: 2026-02-08.md for today's raw notes. What happened, what was built, what broke, what was fixed.
Now every session starts the same way. The agent reads MEMORY.md. It reads yesterday's file. It reads today's file if one exists.
Three files. Maybe five minutes of reading. Then it's caught up.
The difference is subtle but profound.
Instead of explaining context, I reference it. "Remember PBI-45? Same pattern here." The agent knows exactly what I mean because it read the notes from PBI-45 last session.
Instead of repeating mistakes, we compound progress. When something didn't work last week, it's documented. The agent reads it and tries a different approach automatically.
The memory files do something else too. They force clarity.
When I tell the agent to remember something, it writes it down. That act of documentation makes decisions explicit. No fuzzy handwaving. No assuming we're on the same page.
Over time, MEMORY.md becomes a living history of the project. Not a timeline of commits, but a narrative of thinking. Why we chose this architecture. What we learned from that failure. What patterns emerged.
I tried embeddings and retrieval systems first. Fancy vector databases that search through conversation history.
Overcomplicated for what I needed.
Markdown files worked better for me because they're human-readable. I can edit them directly. Add structure as patterns emerge. Remove outdated context when it's no longer relevant.
My agents don't have perfect memory. They have continuity. That turned out to be enough.
If you're fighting the same context battle every session, simple files at session start might help. Worked for me.