Every conversation with an AI assistant starts from zero. Close the session, open a new one, and whatever you explained about your project, your preferences, or your ongoing work is gone. You end up re-explaining the same context over and over, which gets old fast if the assistant is part of your daily workflow. That's the problem Recall is built to solve.
The problem
AI assistants are stateless by default. Whatever "memory" they appear to have is either crammed into the current conversation's context window, or stored by a third-party cloud service you have to trust with everything you tell it. Neither is great: the first doesn't survive a new session, and the second means your notes, preferences, and project details live on someone else's servers.
Recall takes a different approach: a local-first, single-user personal memory system with a CLI and an MCP (Model Context Protocol) server, so an assistant can store and recall persistent context across sessions without any of it leaving your machine.
How it's built
The design is intentionally simple:
- Local-first architecture — there's no cloud dependency for storing personal memory. Everything lives on disk, under your control.
- A CLI for direct interaction — you can read, write, and manage stored memories yourself without going through an assistant at all.
- An MCP server that exposes memory tools (like
rememberandrecall) to any MCP-compatible AI assistant, so the assistant can write new memories and pull relevant ones back in automatically. - Single-user design — no multi-tenant complexity, no auth system to secure, no accounts. It's built to be private and simple, not a hosted product.
It's written in Python, and the whole thing is a relatively small surface area on purpose: a CLI, an MCP server, and a local memory store connecting the two.
What it changes in practice
With Recall wired into an assistant, the loop looks like this: preferences, facts, and project context you'd normally repeat every session get written to local memory once, and pulled back into future conversations automatically. The assistant remembers; you stop repeating yourself.
Because the memory store is local, there's no dependency on a third-party service staying up, no data leaving your device, and no recurring cost. The tradeoff is exactly what you'd expect from a single-user, local-first tool: it's built for one person's machine, not a multi-user deployment.
Try it
Recall is open source. The code, setup instructions, and CLI/MCP usage details are on GitHub. If you're building on MCP or just tired of re-explaining yourself to an assistant every morning, it's worth a look.