Three Ways Claude Can Remember Your Project — And Why I Use Only One

Series: Weekend at Claude's — misadventures in building a production app with an AI anyone could mistake for the person who's going to make the whole thing happen

Claude AI offers three different ways to maintain context across sessions. They sound similar. They work at completely different layers. And choosing the wrong one for the wrong job will cost you in ways that aren't immediately obvious.

This post is specifically about Claude AI — the chat interface. Claude Code has its own session persistence story, which is different enough to deserve its own post. More on that in Vol. 14.

The Problem They're All Solving

Claude AI has no memory between sessions. Every conversation starts fresh. For a casual use case, that's fine. For a production app with architectural decisions, dependency-ordered build tasks, legal blockers, and a fundraising timeline, it's a real problem.

Starting a session without giving Claude context means twenty minutes of re-orientation before you can do anything useful. Getting Claude oriented is the job. The question is how.

Option 1: Project Knowledge Files

I maintain a set of structured markdown files — TASKS.md, CONTEXT.md, DECISIONS.md, HISTORY.md — and upload them at the start of every session. Claude reads them, orients itself in about thirty seconds, and we get to work.

Key properties:

You control exactly what Claude knows. The content is what you wrote. No inference, no automated extraction, no surprises. If a decision changed, you update the file.

It's structured and queryable. TASKS.md has status columns, blocker columns, dependency order. Claude can reason about that structure — it can tell you what's ready, what's blocked, what the highest-value next action is. Loose notes can't do that.

It works everywhere. These files live in your git repo. Claude Code reads them. Other tools can read them. If you change tools, your context comes with you.

It requires maintenance. The honest cost. You have to update the files; they don't update themselves. For a project with real business stakes, that discipline is a feature — it forces a clear record of where things actually stand.

Option 2: Search / Reference Past Chats

This lets Claude search your previous Claude AI conversation transcripts. "What did we decide about X last month?" and Claude goes looking.

Useful as a fallback — a search index for things you didn't think to write down. Not a replacement for structured context:

Transcripts are noisy. Claude has to find signal in thousands of lines of conversation and might miss things or surface the wrong session.

It finds what was said, not what was decided and still stands. A reversed decision might surface without its reversal.

Only works in Claude AI — invisible to Claude Code and everything else.

Think of it as a useful supplement. Not a foundation.

Option 3: Auto Memory from Chat History

Claude automatically extracts facts it thinks are worth remembering — your name, your stack, your preferences — and surfaces them across future conversations.

Useful for basic personal context. Not useful for project management:

You don't control what gets captured. Claude decides what seems memorable.

No structure. Memories are loose facts, not a dependency graph.

Can go stale without you knowing. A reversed decision might still be remembered as current.

Lives only in Claude AI — invisible everywhere else.

How They Compare

What I Actually Do

I maintain the project files. That's the primary mechanism, and it's non-negotiable for anything serious.

I also have past chat search and auto memory enabled — they don't conflict with the files, and occasionally one of them surfaces something useful. But they're fallbacks, not foundations.

The deeper principle: for anything with real stakes, you want context you wrote, in a format you can reason about, stored somewhere that travels with the project. Automated extraction is convenient. Deliberate documentation is reliable.

In the movie, the guys had to physically manage Bernie's presence at every location, every encounter, every situation. Nobody was doing it for them automatically. It was exhausting. It also worked.

Write it down yourself. Update it before each session. It takes five minutes and saves twenty.

The Actual System: Two Checklists

Here's exactly what I do. Copy these if they're useful.

Session Kickoff

1. Upload these files (drag-and-drop, all at once): - TASKS.md — always - CONTEXT.md — always - DECISIONS.md — if decisions were made since last session - aphilaty_architecture.md — only if architecture changed 2. Paste this prompt: "Starting a new Aphilaty session. Uploading current project files now. Please wait for all uploads before responding." 3. After uploads confirm, say: "Session kickoff — please review TASKS.md and tell me where we are and what to work on."

Claude reads the files, confirms status, flags blockers, recommends the highest-value task. Re-orientation: thirty seconds.

Session Close

Before closing the session:

1. Download all output files generated this session 2. Paste this prompt: "Session wrap: summarize what we accomplished, decisions made, any new gotchas to add to CLAUDE.md, and updated task statuses. I'll use this to update HISTORY.md, DECISIONS.md, and TASKS.md." 3. Update TASKS.md, DECISIONS.md, and HISTORY.md from the summary 4. Commit downloaded outputs to the repo 5. Upload updated TASKS.md and CONTEXT.md to the Claude.ai project knowledge so they're current for next session

The close checklist is the one that's easiest to skip when you're tired. Don't. The five minutes you spend closing a session properly is the thirty seconds you save at the next kickoff — multiplied by every session that follows.

For how Claude Code handles this differently — with custom /session-start and /session-wrap commands Claude built for itself — see Vol. 14.

Aphilaty is a privacy-first community coordination app. aphilaty.com