Building an External Memory So Claude Stops Forgetting Everything on Reset

Losing an entire conversation's worth of context when a session maxed out felt like starting over from nothing. Not annoying. Actually costly. Hours of decisions, half-formed plans, the specific reasoning behind a choice I'd made two days earlier, gone the moment the context window filled up.

That specific pain led me down two connected questions, and I want to walk through both, because the answer to the second one changed how I build with AI in a lasting way.

The first question: what tone actually helps an ADHD brain retain information from AI. I went looking for a real answer instead of a guess. What I found: lead with the point, not the windup. One idea per message, not five bundled together. Concrete over abstract. Warm over neutral. A flat, neutral tone is forgettable, and guilt-based language, the kind that nudges you by making you feel behind, actively backfires. It doesn't motivate. It makes people disengage.

The second question was the bigger one, and it's the one this article is really about: how do you stop AI from forgetting everything on a session reset.

The instinct is to want a bigger context window. More room, more memory, problem solved. That's not where the actual answer was. The answer was building external memory on purpose, as its own deliberate system, instead of hoping the model's internal context would carry it.

The pattern is simple to state: read at the start, write at the end. Every session opens by reading the current-state page. Every session closes by updating it, and adding one entry to the log.

The actual two-file system

### current-state.md (always overwritten, never appended to)
- What's being built right now
- Key decisions made and why
- Open questions still unresolved
Read this first, every session.

### session-log.md (append-only, one entry per session)
## [Date] — [one-line session summary]
- What happened
- What changed
- What's next
Read the last 1-2 entries when picking up a stalled thread.

Here's the actual lesson, and it's the one I keep coming back to: the fix for limited working memory isn't a bigger container. It's reducing what has to live inside it at any given moment. That's true whether the working memory in question belongs to a model resetting its context, or a person trying to hold too many open threads in their head at once. A two-tier system, one page for current state, one log for history, solves the same problem in both cases.

If you're running long AI-assisted projects and losing context between sessions, don't wait for a bigger window. Build the two files. One overwritten, one appended. Read at the start, write at the end. It's a small system, and it's the reason a session reset doesn't cost me the last two weeks of thinking anymore.