Launching today

Bastra Recall

Launching today

Your AI's working memory. Local, open, in Markdown.

2 followers

Your AI's working memory. Local, open, in Markdown.

2 followers

•

•

Every new AI session starts from nothing. Bastra Recall gives your AI tools a memory that stays: preferences, decisions and hard-won fixes live as Markdown files on your own machine, ready again next session. Works with Claude Code, Claude Desktop, Codex/ChatGPT Desktop and Cursor. The files are yours: open them in Obsidian or any editor. Nothing leaves your machine, no account, no cloud. Open source under MIT. macOS fully supported, Linux runs daemon, CLI, MCP and hooks. Windows in progress.

Free

Launch tags:Open Source•Developer Tools•Artificial Intelligence

Launch Team / Built With

Agent Builder by AirtopWeb agents that heal themselves.

Promoted

Maker

📌

Hi Product Hunt, I built Bastra Recall because of something that annoyed me every single day: I keep explaining the same things to my AI assistant. What we decided last month, why this particular detour is necessary, how I like things built. Every new session starts from nothing. There are cloud memories for this. What mattered to me was that my working notes do not sit on someone else's server, and that I can read them without asking permission. So this is how Recall is built: Your memories are ordinary Markdown files in a folder you choose. You can open them in Obsidian, edit them by hand, or delete them. Nothing leaves your machine, and there is no account. Recall happens on its own. A local service hooks into the places where your assistant works and puts the relevant memory in front of it before anything gets built. You do not have to remember to ask for it. The search understands what you mean, not just which words you typed. Ask about rounding and money and you get the decision about integer minor units, even though not one of your words appears in it. The model behind that runs locally too. And you can check what it costs you. Version 1.0 added a telemetry view that shows how many tokens Recall writes into your context and how often a surfaced memory was actually used. From your own logs, computed locally. Version 1.0 carries ninety-one closed issues, eight rounds of counter-review by an independent agent whose only job was to find what was broken, and 2,910 tests. On platforms, plainly: macOS is fully supported, including guided setup, autostart and the Claude Desktop extension. Linux is further along than the short version suggests. The daemon, the CLI, the MCP surface and the compiled hook client all run there, the release carries binaries for x86_64 and arm64, and our test runs pass on Ubuntu. What is missing are two conveniences: installation goes through npm rather than Homebrew, and there is no autostart equivalent yet, so the service starts on demand and shuts down after thirty minutes idle. Both are tracked as issues. Windows is not there yet, and I am not going to talk around it. What it needs is written down one by one: paths in Windows convention, a service manager instead of the macOS mechanism, the compiled hook client, and the path for documents. Full coverage of all three systems is the goal we keep working toward. If you are on Linux or Windows and need one of those badly, say so here. That is the best way to influence the order. Open source under MIT. Send me your questions, I will answer as fast as I can. And tell me: what is the one thing your AI assistant keeps forgetting?

Report

19h ago

Forum Threads

p/bastra-recall

•

18h ago

What does your AI assistant keep forgetting, even after you have explained it three times?

I am building a memory for AI assistants and I would rather collect real cases than assumptions before I go further.

For me it is three kinds that keep coming back. Decisions first: why we took a detour two months ago that every newcomer immediately wants to straighten out again. Then preferences: how I want things written, explained from scratch every time. And the expensive ones, hard-won fixes where I know I have had this problem before but no longer how I solved it.

What I am actually curious about: have you built a habit against it? I mostly see three approaches. Some keep a rules file in the project and stop maintaining it after a while. Some paste a block of context at the start of every session. Some have made their peace with it and just explain again.

And the question behind it: where would you want that memory to live? On your machine, in files you can read, or with a service that takes the work off your hands and keeps the data? I went with the first, but I am genuinely curious how many of you see it differently.

View all