ij5a/subrosa

README preview

<p align="center">
  <img src="assets/subrosa-wordmark-animated.svg" alt="subrosa" width="100%">
</p>

<p align="center">
  Persistent, private memory for Claude Code — that never spends your tokens on itself.
</p>

<p align="center">
  <a href="https://github.com/ij5a/subrosa/releases/latest"><img src="https://img.shields.io/github/v/release/ij5a/subrosa" alt="latest release"></a>
  <a href="https://github.com/ij5a/subrosa/actions/workflows/ci.yml"><img src="https://img.shields.io/github/actions/workflow/status/ij5a/subrosa/ci.yml?branch=main" alt="CI status"></a>
  <a href="LICENSE"><img src="https://img.shields.io/github/license/ij5a/subrosa" alt="MIT license"></a>
</p>

Every session is archived into a local SQLite database and made searchable: last month's work is one `subrosa search` away, relevant past sessions resurface when you type a related prompt, and Claude stops rediscovering what it already figured out. Pulling up an old answer costs a few hundred tokens; having Claude re-derive it from scratch runs into the thousands.

- **No LLM calls to save memory.** Saving a session is plain-text parsing — zero tokens. Most memory plugins run your sessions through an LLM to save them ([the comparison](docs/comparison.md) has the numbers, from their own docs).
- **Hard token limits, set in the code.** Recall adds ~180 tokens on a strong match, usually nothing — it stays silent otherwise. The always-loaded index is capped at 23 KB. [Check it yourself](#proof-verify-it-yourself), or see [where your tokens go](docs/faq.md#how-many-tokens-does-it-cost-me).
- **One ~4 MB static binary.** No daemon, no worker port, no background process — a hook fires, finishes in under 10 ms, and exits.
- **Your transcripts stay on your machine.** The binary makes zero network calls — no cloud, no telemetry — and obvious secret shapes are masked before storage. [Verify every claim yourself](#proof-verify-it-yourself).

View full repository on GitHub →