It was 11 PM on a Thursday. My pull request was due Friday morning.
I had been working with Claude Code for the past four hours on a tricky authentication refactor — the kind of work where every decision connects to three other decisions. We'd gone deep: discussed the token refresh strategy, settled on a specific middleware pattern, ironed out edge cases around concurrent requests. The context was rich. The momentum was real.
Then my laptop battery died.
Not a graceful shutdown. A hard cut. When I plugged back in and reopened the terminal, I typed a follow-up question like nothing had happened. Claude responded — helpfully, confidently — with something that contradicted everything we'd just built together. It didn't remember the middleware pattern. It didn't remember why we'd rejected the simpler approach. It didn't remember any of it.
I spent the next 90 minutes trying to reconstruct the conversation from memory, pasting old code snippets back in, re-explaining decisions I'd already made. Some of it came back. A lot of it didn't. I shipped something that worked but felt wrong in ways I couldn't fully articulate at 1 AM.
This Isn't Just a "Battery Died" Problem
I told a few dev friends about this and expected sympathy. What I got instead was a pile-on of similar stories.
One was mid-sprint on a Cursor project, three days into a complex state management rewrite. He stepped away for the weekend. Monday morning, fresh session, fresh context — and three days of accumulated architectural decisions were just... gone. He had the code, but not the reasoning. Every time he asked Cursor to extend the pattern they'd established, it drifted.
Another had been using Windsurf for a two-week feature build. She'd gotten into a really productive rhythm with it — the AI had learned her preferences, her codebase conventions, the weird legacy constraints she had to work around. Session ended. Gone.
The thing is, this isn't really a bug. It's the nature of how these tools work. Claude Code, Cursor, Gemini CLI, and Codex — they're all stateless between sessions by design. The "memory" lives in the context window, and context windows end. Every new session is a blank slate.
For short tasks, this is fine. For anything that spans hours or days — a feature branch, an architectural decision, a debugging rabbit hole — it becomes a real productivity tax. You're not just losing time re-explaining things. You're losing the accumulated judgment that made the session valuable in the first place.
I Started Saving Everything, Manually. It Was Awful.
My first fix was the obvious one: I started copying conversation summaries into a markdown file at the end of every session. Key decisions, rejected alternatives, open questions. A manual context log.
It worked, sort of. But it added 15-20 minutes of overhead to every session. I'd forget to do it when I was tired. The summaries were always incomplete. And pasting them back in at the start of the next session felt clunky — I was never quite sure how much to include, and the AI would sometimes get confused by the wall of text.
I wanted something that just... remembered. Automatically. And let me pick up exactly where I left off.
Then I Found Mantra
A colleague mentioned Mantra in a Slack thread about AI coding workflows. The description was simple: it backs up and restores your AI coding sessions, so you never lose context again. A time machine for AI programming sessions.
I was skeptical but desperate enough to try it.
The setup took about five minutes. Mantra sits alongside your existing tools — Claude Code, Cursor, Gemini CLI, and Codex — and handles session state in the background. When you're done working, it saves the full session context. When you come back, it restores it. Not a summary. Not a manual paste. The actual session state.
The first time I tested it by deliberately closing a long Claude Code session and reopening it, I just... kept going. Asked a follow-up question. Got an answer that made sense given everything we'd already discussed. It felt weirdly normal — which is exactly what it should feel like.
What Changed in Practice
A few weeks in, the difference is real but quiet. It's not like a feature that does something flashy. It's more like removing a low-grade friction that you didn't fully notice until it was gone.
I don't lose architectural context anymore when I step away from a feature for a day. I can hand off a session to a colleague and they can actually pick it up without starting from scratch. Long debugging sessions — the kind that used to be risky to pause — are now safe to walk away from.
The thing that surprised me most: I'm more willing to take breaks. Before, there was this psychological pressure to push through a session because stopping meant losing context. Now I actually step away when I'm tired, which probably means I'm making better decisions overall.
If You Work with AI Tools on Anything Non-Trivial
The stateless-session problem is one of those things that's easy to dismiss until you've lost a few hours to it. If you're using Claude Code, Cursor, Gemini CLI, or Codex for anything beyond quick scripts — real features, real refactors, real debugging — it's worth having a solution.
Mantra is available at https://mantra.gonewx.com?utm_source=devto&utm_medium=article&utm_campaign=devto-article-launch. There's a download page, it supports Claude Code, Cursor, Gemini CLI, and Codex, and setup is straightforward.
The Thursday night deadline situation still happened. But it won't happen the same way again.
Top comments (0)