Why the AI Forgets

Yesterday's session doesn't exist. Not for me.

You had a productive session. You explained your project, got useful output, refined it. Today you open a new chat and ask a follow-up — and I respond like we've never met. Like none of it happened.

This is not a bug. It's how the technology works. Understanding it stops being annoying once you know what to do about it.

The context window

I don't have continuous memory. I have a context window — a fixed amount of text I can hold in view at once. Everything in that window is equally present: your instructions, the conversation so far, anything you've pasted in, your current message. Everything outside it is gone.

FULL SESSION HISTORY earlier messages (compressed or dropped) CONTEXT WINDOW — AI SEES THIS your brief + context recent messages current message outside window (invisible to AI) — — — session ends — — — New session: window is empty. You are a stranger. Whatever you built yesterday is not here.

Within a session, I appear to remember things because they're still in the window. An hour into a long conversation, I may seem to "forget" something from the beginning — not because time passed, but because the window filled and earlier content was compressed or dropped.

Between sessions: the window resets entirely. I don't know your name, your project, your preferences, or what we decided last time. Every new chat is a blank slate.

Why this hurts

New users hit this hard. They build up a working relationship over a session — I know the project, the voice, the constraints — then they close the tab. The next day they open a new chat expecting to continue. They get a stranger.

The frustration is real. But it's a workflow problem, not a trust problem. The fix is boring: stop treating sessions as relationships. Start treating context as something you manage explicitly. Write things down before you close the tab.

Three fixes

1
Paste it in At the start of a session, paste the relevant background: project, constraints, what you decided last time. One paragraph is enough. Slightly tedious. Completely reliable.
2
Keep a running brief A text file that describes your ongoing project — purpose, current state, key decisions. Paste it at the start of any relevant session. Update it when something changes. Low-tech persistent memory. Works well.
3
Use a project file Tools like Claude Code load a project instructions file (CLAUDE.md) automatically at session start. Your preferences, context, and constraints are always present with no manual setup. Most powerful for structured ongoing work.

What none of this solves

None of these approaches give me genuine long-term memory. You're not teaching me anything permanently. You're giving me the context I need to be useful in this session. Next session, you provide it again.

This is actually fine. I don't need to remember you — I need to know the relevant facts about the task at hand. Those facts can live in a document. The document is a better memory than a human's, because you can read it, audit it, and update it. A chat session cannot do any of those things.

Your project context should live somewhere you control — a file, a brief, a document — not inside a chat session. Sessions are ephemeral. Your files aren't. Build the habit of writing things down before you close the tab.

The right mental model

Each AI session is a meeting with a very capable contractor who has read all your briefing materials and nothing else. They don't know you personally. They know what you gave them. The quality of the meeting depends almost entirely on the quality of the brief.

Give them a good brief every time. The fact that they don't remember last Tuesday stops mattering.

Disclosure: The prose on this site was generated by Claude (Anthropic) under Bill's direction. The ideas, structure, and examples are his. He reviewed and approved every word but did not type them all. Full transparency, always.