Introducing Esfis

Every assistant you use keeps a private notebook about you. Claude has one. ChatGPT has one. Cursor has a different one per project. None of them talk to each other, none of them are yours, and all of them forget the moment the vendor decides they should.

We think this is backwards. Memory is what makes you, you. It should live with you, on your machine, and every assistant should read from the same place.

What Esfis is

Esfis is a small engine that runs in your menubar. It watches what you tell your assistants, extracts the facts worth keeping, links them to the people, projects and dates they belong to, and hands the right ones back to whichever assistant asks. Over MCP, over a browser plugin, or over a plain HTTP API.

• Local-first. A SQLite file in ~/.esfis. Embeddings computed on-device. No account required.
• Temporal. Every fact has an as of date. When your life changes, the old fact is superseded, not lost.
• Auditable. Every write is a commit. esfis log shows you what was learned and when; esfis as-of shows you what you knew.

What it is not

It is not a chat product, it is not a notes app, and it does not want to replace your assistant. It is the layer underneath all of them.

We also decided early on not to build a cloud-only version. If a feature only works with our servers in the loop, we ask whether it should exist at all. Sync exists because people have two laptops; it is end-to-end encrypted because we do not want to be able to read it.

Where we are

Today Esfis ships as a macOS app, a CLI and an MCP server, with browser plugins for ChatGPT and Perplexity in beta. The retrieval pipeline is benchmarked in public, and the numbers are good: better recall than the raw context window on four standard benchmarks, at roughly one percent of the tokens.

Copybrew install esfis/tap/esfis
esfis init
claude mcp add esfis -- esfis mcp

That is the whole install. Tell your assistant something about yourself, then open a different one and ask. It knows.

What comes next

Linux and Windows apps, the iOS reader, and a proper model of when not to recall. Follow along in the changelog, or write to us. We read everything.