For Claude Code users, the most annoying part of a long-running project is not the coding itself — it’s the constant re-introduction. Recall is interesting because it tackles that cold-start problem with a very opinionated approach: local-first, offline, and deliberately non-LLM. That makes it less flashy than “AI memory” products, but in some ways more practical.
.recall/ in your project:
history.md: an append-only session log.context.md: a compact summary meant to be loaded in later sessions.pip install, no local model setup, and it works offline.CLAUDE.md is manual instruction memory.--continue / --resume replay prior conversations.history.mdcontext.md/recall:save regenerates the summarycontext.md includes the summary plus deterministic facts such as the goal, files touched, commands run, where you left off, and git diff stats.recall.config.json, with options for output location, history capture, auto-save, redaction, Git metadata, and input size limits..recall/.capture-paused, to stop logging without editing config.
What strikes me is that this is not trying to be “AI memory” in the usual overpromised sense. It’s more grounded: capture the session, summarize it locally, and make the result easy to reopen later. I think that’s a much more credible product shape for Claude Code than a cloud memory service that quietly sends your repo context somewhere else.
I also like the constraint of using extractive summarization instead of another model call. That feels almost old-school, but in a good way: predictable, cheap, and aligned with the privacy story. If the summary quality is decent, this might be exactly the sort of boring infrastructure people end up relying on every day.
What I’d be curious about is how well it holds up on messy, nonlinear sessions. TF-IDF and TextRank can do a respectable job at surfacing central sentences, but coding work is often more about decisions, failed branches, and “we tried that already” context than about sentence centrality. My guess is that the deterministic metadata helps a lot, but I’d still want to test it on real projects before trusting it fully.
The privacy angle is the strongest part of the pitch. If you’re working in a sensitive repo, the idea of keeping transcripts local and fenced inside the project is genuinely useful, not just marketing. The only caveat is that any tool writing session history to disk becomes something you should understand and audit, especially if secrets can appear in prompts.
As a Claude Code user, I’d actually try this on a real project where I keep coming back after interruptions. That is exactly the workflow pain it targets, and it seems thoughtfully built around Claude Code’s existing memory features rather than trying to replace them.