Surfing The Singularity home

How Claude remembers things, for people who don't code

What loads every time, what loads on demand, and what never loads at all.

Surfing The Singularity

This page expires. Published September 21, 2026. Assume it is obsolete after December 20, 2026. If you are reading this more than 90 days after it went up, don't spend your time here. The products change too fast for these specifics to last.

What's ahead: about eight minutes. Verified facts, rules of thumb, and one forecast sit in separate sections, and each one is labeled. This is the companion to "Why are there CLAUDE.md files on my computer?"

The short version

Claude does not remember you between conversations on its own. What looks like memory is text that gets loaded in. Some text loads at the start of every conversation. Some loads only when something asks for it. Some never loads at all, even though it sits right there on your disk. Knowing which is which saves you usage, time, and confusion.

The Claude app also has optional memory features in its Settings. Those are outside this piece.

Two tiers in the Claude app

Fact

Sources for this section: Anthropic's support article on usage limits and a third-party guide from likeone.ai. Both were read September 21, 2026.

Tier one — every conversationTier one loads on every single conversation. That is your Project instructions, the text box inside a Project. It is added before every chat in that Project, including the trivial ones. A one-line question still carries the whole instruction block with it.

Tier two — on demandTier two loads only when needed. Files uploaded to a Project are retrieved when they're relevant. Files in a folder you connected from your computer load only when something reads them.

It is easy to get this backwards. You can pour hours into a big instruction file on your drive that never loads. Meanwhile the Project instructions box costs something on every chat. The text box you type into once is the one you pay for every time.

How Claude Code does it differently

Fact

Source: Anthropic's Claude Code memory documentation, read September 21, 2026.

Claude Code loads CLAUDE.md files automatically, based on where they sit. It reads CLAUDE.md from the folder you start in and every folder above it. Files in subfolders load when Claude opens files in those subfolders. Claude Code also keeps notes for itself in a separate auto memory folder.

The phrase "Project instructions" means three things. In the Claude app, it is the text box inside a Project. In the Claude Code docs, it is the label for a folder's CLAUDE.md. In Claude Code's settings panel, it is also the name of a setting that picks which instruction files load. The same two words do three jobs.

Rules of thumb (not official figures)

Rule of thumb

Tokens are how Claude measures text, and they drive your usage. Here is a working rule. Dense markdown runs about 3.5 characters per token, and plain prose runs about 4. So divide a file's size in bytes by 3.5 for a rough count. A 27KB file comes to roughly 7,000 to 8,000 tokens. These are estimates for reasoning, not Anthropic numbers.

If that file sits in your Project instructions, every chat carries those tokens. If it sits on your drive and nothing reads it, it costs nothing and does nothing.

What the research says about file size

Fact

Anthropic gives a number and a test. The Claude Code memory page recommends keeping each CLAUDE.md under 200 lines. It says longer files use more context and reduce how well Claude follows them. Some secondary sources report this figure, and I confirmed it on Anthropic's page on September 21, 2026. The best-practices page adds a test for every line: "Would removing this cause Claude to make mistakes? If not, cut it."

Anthropic also says Claude Code loads a CLAUDE.md up to 4 MiB in full and skips anything larger. The linting site claudelint.com reports a Claude Code warning at 40KB. I have not confirmed that warning on Anthropic's own pages.

One study found no effect on obedience. As reported by alexdunlop.com, one test compared files from 25 to 500 lines. It found no measurable difference in how well Claude followed the rules (χ²(3)=5.16, p=0.16). A separate ETH Zurich evaluation, reported on the same site, found context files raised costs by over 20% on average. They did not raise success rates.

So the solid case for short files is cost. Anthropic says long files hurt adherence. One independent test found no measurable effect. The cost shows up either way, because every loaded line uses tokens. Keep files short for that reason, whether or not length changes behavior.

What to do with this

A forecast, clearly labeled

Forecast

This section is my opinion about the future, not a fact. Neither Claude nor I have any knowledge of Anthropic's plans. My guess is that Claude's separate products will merge into one. It reminds me of phones. A phone used to be a phone, and a camera was a camera. A map was a paper map in the glovebox. Then one day it was all just the phone.

One fact points that way. Cowork was once a separate Anthropic product, and its features have since folded into the Claude app. Anthropic announced the merge on September 16, 2026. Cowork is now Claude. If the merge continues, the CLAUDE.md confusion may disappear on its own. It may also get worse before it gets better.

So which tier has your effort been going into?

Sources

All read September 21, 2026.

Come find me on Substack

This website is an archive. My Substack is where I write to my son, and it's the only place I answer messages. If something here landed, that's where to reach me.

Surfing The Singularity on Substack