with a memory, so every session doesn't start from zero?

The problem

Every AI chat starts blank. For a while I spent the first part of every session re-explaining myself: who I am, which project this was, what we'd decided last week, what not to touch. The AI was good. It just had no yesterday. The more I used it, the more of my time went into catching it up.

What I actually run

One folder that Claude Code reads first, every session. Everything in it is a plain text file I can open and edit myself.

• A page about me and how I work. Read at the start of every session, so I never introduce myself again.
• A list of every project, each with what phase it's in, when it last moved and the one next step. It's close to fifty projects now, from client video work to the lights in my apartment.
• Lessons from what broke. When something fails and gets fixed, the reason goes in a lessons file, so the same mistake isn't solved twice.
• A wrap-up at the end of a session. It files what happened today so tomorrow's session starts where this one stopped.
• A weekly check that compares the records against what's really there, and flags anything that has drifted.

It runs on Claude Code and a folder of text files, backed up with git, and the same folder is shared across the three computers I work on.

What it took

It didn't arrive finished. It grew a piece at a time over months, and each piece exists because something went wrong without it. Running it needs a paid Claude plan and nothing else: no extra service, no database. You don't need to start where I am. The free starter below is the first afternoon of it.

What broke

• A backup was dead for 32 days and its log said so the whole time. Nothing was reading the log. Now everything that runs on a schedule writes one status line per run, and something reads those lines. A log nobody reads is just a slower silence.
• "Next session should check this" sat untouched for 23 days. A note in a file only gets read when something happens to open that file. Anything that has to happen later now gets a date that the weekly check reads, or it gets done now.
• The summary went stale and was trusted over the real list. My top-level page said three projects were shared between computers when six were, and a session believed the page. Now the page says which file is the authority, and the summary is only a signpost to it.
• The first person to try my starter set it up perfectly and still asked "okay, now what?" Every part worked. Nothing told them what the folder was for. That gap is why this library exists: each entry is one real answer to "now what?"