Bastra Recall
Your AI's working memory. Local, open, in Markdown.
5 followers
Your AI's working memory. Local, open, in Markdown.
5 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 Team / Built With
Dropbox Sign APISimplify your eSignature integration
Promoted
@n0mad
Neat launch.
Building a similar memory layer on an agent harness. Recall quality isn't the hard part — the trigger is. Fire every turn, it's noise. Fire only on explicit "remember", you miss facts dropped mid-sentence without a flag. How does Bastra decide that boundary: similarity threshold, or checking the fact is actually missing from context first?
Second: how are notes typed internally? I split soft reminders from hard constraints ('"never X, because Y") from recommendations a model can override when circumstances differ from the one that produced the rule.
Does that hold across domains, or does code — which has a live source of truth — need different trigger logic than something
without one?
@ivan_bush Thanks, Ivan! You hit the sore spot.
On the trigger: there isn't one trigger, there are several points, each with its own rules.
- Before a file edit, a hook builds a query from the path, the file extension and the content, and surfaces matching memories.
- On the prompt, Recall does not search every turn. It only fires when the prompt asks for memory, or asks for outbound text or a statement about the project's state.
- Individual memories can be marked as reflex. Those fire deterministically when every word of one of their triggers appears in the prompt, at most two per turn.
Against noise: score floors, every memory shows up only once per session, and a source whose hints the agent keeps ignoring backs off. Whether a fact is already in context by content, I honestly don't check. I only check whether Recall has already shown it in this session. That's the open half.
Saving works the same way: not only "remember this" counts, but also rules, corrections, decisions and frustration. At the end of a turn, a hook reads the last messages and suggests saves, but it never saves on its own. Facts dropped in passing without any signal still slip through today.
On types: lesson, decision, preference, workflow, project-fact and a few more. There's no explicit hard/soft split like yours. Instead, every memory carries a "why" and a "how to apply" in its body, so the model can see whether the circumstances still match. New versions explicitly replace old ones, and contradictions are flagged as a visible conflict instead of being silently overwritten.
And yes, code works differently. There the trigger sits right on the edit tool, and a project fact can carry a verification command that checks the claim against the current code. It never runs automatically; the agent decides. Without such a source of truth, all that's left is age, usage and explicit replacement.
I am curious how you check whether a fact is really missing from context. That's the piece I'm still missing.