My AI Was Running on Half Its Memory — and Never Said a Word
My AI assistant had incomplete memory for weeks. I had no idea.
OpenClaw boots with context files — project notes, past decisions, infrastructure details. My main memory file had grown to about 25,000 characters over months of daily use. There’s a config setting that caps how much of each file loads at boot. Default value: 12,000 characters. So every session was starting with the bottom half of my notes silently missing.
The AI never flagged it. Answers were coherent and confident the whole time. It just worked with whatever portion had loaded, as if the rest didn’t exist.
The only clue: a bootContext failed error that appeared in the system feed one evening. Traced it in about a minute — file exceeded the limit, bottom half never parsed. One config change (12,000 → 30,000) and the full file’s been loading since.
I’d been building that memory system for months: project notes, infrastructure details, lessons from past mistakes, ongoing thread summaries. The file probably crossed the 12,000 character limit about two months ago. Since then, every session started with half the notes missing.
If you’re running any AI tool with persistent memory — your own server, a third-party service, anything that carries state across sessions — check the limits. The system won’t warn you when it’s hitting them. Confident answers from half the data look exactly like confident answers from all of it.
OpenClaw is a self-hosted AI assistant that runs on a server and handles real daily work: email triage, publishing, research, automation. Memory across sessions is one of its core features; apparently worth auditing occasionally.
I teach a class on setting up and getting the most from OpenClaw — details at themeperks.com/openclaw-course/.
