How are you handling the "cold start" problem in Claude Code today ?

Hey Product Hunt 👋 I'm the maker of Recall (github.com/raiyanyahya/recall).

The itch that started it: Claude Code begins every session cold. I was re-explaining my project — goal, where I left off, what I'd already tried — at the start of almost every session, and it felt like a tax I kept paying. So I built Recall to fix that without the thing I didn't want: shipping my transcripts off to some endpoint.

How it works, briefly: a hook logs each session into .recall/history.md as you work, and a local summarizer condenses it into a compact context.md (~1–2K tokens) you load into the next session — goal, files touched, commands run, next steps, where you stopped. The deliberate constraint is that there's no LLM call anywhere. The summary is built by a classical TF-IDF + TextRank summarizer running entirely on your machine. No API key, no external model, works offline, and it spends zero model tokens so persistent memory costs nothing beyond the subscription you already pay for.

Two design decisions I made on purpose, and I'm genuinely curious whether you'd have made them differently:

1. Fully local, even at the cost of "smarter" summaries. Most memory tools pipe your context (code, paths, sometimes secrets) to a model to generate a nicer summary. I chose a deterministic local algorithm instead, so the privacy guarantee is structural, nothing can leak because nothing is ever sent. Do you feel that trade-off is worth it, or would you take a smarter abstractive summary if it stayed on a trusted endpoint?

2. Extractive over abstractive. TextRank picks the most central sentences rather than rewriting them. In my own use the digest resumes cleanly, but I'd love to hear from anyone who's tried it: does it capture where you left off well enough, or do you find yourself topping it up?

A couple of things I'm still chewing on and would love input on:

- Manual /recall:save vs. the auto_save_context: "on_end" option that regenerates on every session end — which do you reach for?

- How are you handling the cold-start problem today? CLAUDE.md, --resume, copy-paste, or just eating the re-explanation each time?

For the record, I see Recall as complementary to Claude Code's built-ins CLAUDE.md is how I want you to work, --resume replays a full transcript, and Recall is an automatic, deterministic record of what each session did and where it stopped, produced offline.

It's MIT licensed and on GitHub . Happy to answer anything about the internals, the summarizer, or the privacy model. What would make this genuinely useful in your workflow?