> Source: https://www.esfis.com/en/blog/context-not-memory · esfis, your projects, your context, your agents. Updated 17 September 2026.

# Context, not memory

> Remembering across sessions is something the assistants already do. What none of them does is decide which context each AI should receive, where it came from and whether it still holds. So esfis is changing category.

Francesc Pinyol, Founder · 14 September 2026 · 2 min read

When we introduced esfis a week ago, we summed it up in one line: *a memory for every assistant*. It was clear and it was true. It was not the right line.

## Remembering is no longer the problem

Claude Code has `CLAUDE.md` and automatic memory. Codex has `AGENTS.md` and memories. Cursor has its rules. With every month that passes, remembering things between sessions looks more like a platform feature than a product, and there is already a healthy crowd of tools selling memory for AI, some of them very good.

If esfis were one more memory, it would be competing with all of those tools and with the vendors themselves. And it would lose.

## The question still open

The question no assistant answers on its own is not *how do we get every AI to remember the same things*. It is this one:

> What should each AI know, right now, about this project, and why?

That question will still be there on the day every vendor's memory is perfect, because none of them is yours and none of them talks to the others.

## Three things esfis knows

**Where it came from.** Every piece of context records its source, its date and whether it was stated outright or inferred. An inference is never presented as a fact, and it never goes into a project file.

**When it changed.** A new decision supersedes the old one, which becomes history rather than living alongside it. And every item can carry an expiry date.

**Who can see it.** Projects and clients have spaces of their own, and what belongs to one client never reaches another: the space is checked before anything is searched, not after. You decide which AI sees which space, and `esfis --why` explains what each one received and why.

## Keep your CLAUDE.md

We do not want you to abandon your instruction files. `CLAUDE.md` tells Claude how to work; esfis decides what context it should have. They are two different things, and esfis can write the second one inside the first:

```sh
esfis --compile claude --space esfis
```

Inside a delimited block, only what is current, explicit and permitted goes in. What you wrote yourself is left alone, and compiling again with no changes does not alter a single byte.

## What we will not do

We will not record the screen or read all your mail just in case. *The minimum useful context* is the opposite of *all observable activity*: a memory full of noise is worse than a small one, and far harder to protect.

## How we will know it is working

We will not count how many items get stored. We will count how often one AI makes use of context learned by another without you having to explain it again. And we will watch one number that must always be zero: context from one space reaching another.

All of this is available in [version 1.5](/en/changelog) and in the [documentation](/en/docs#context).