Hydrate in ten minutes.

Hydrate is an episodic-memory layer for Claude Code. The shape of it: every time you write a prompt, the facts you need from previous sessions surface in the model's context before the prompt is sent. You did not ask for them; they were already there. The capture side is symmetric; every session you finish is read for new facts and stored locally. Nothing to think about after the install.

The shape of it

1. You run curl -fsSL gethydrate.dev/install | sh. The installer drops five binaries into ~/.local/bin/, wires two hooks into ~/.claude/settings.json, and starts hydrate-server in the background.
2. hydrate-server runs on a local port (written to ~/.hydrate/server.port; default 49849) - it owns the SQLite store and serves the dashboard.
3. Every time you run claude, the UserPromptSubmit hook queries the server and injects relevant facts. When the session ends, the Stop hook saves a summary and extracts new facts.
4. Run hydrate doctor to verify everything is wired correctly.

Next steps

• Install - the one-line setup + 15-check doctor
• Quickstart - verify it's working
• Builders - generate Claude Code skills, Hydrate hooks, and phased plans from a one-line prompt. The fastest way to make Hydrate productive on day one.
• BYOK - bring your own keys - how to add provider API keys, the right way
• Team - git-as-sync-layer for shared project memory
• Scopes - how facts stay organised
• Enterprise - self-host + SSO + org policy

Core concepts

• Pre-prompt injection: the UserPromptSubmit hook queries the local daemon and prepends relevant project context to every prompt before it reaches the model. Retrieval runs on FTS5 keyword match plus optional cosine similarity, with a recency-decay weight.
• Team Memory as Git: a project's canon and facts live in a git repo you control. Push from one machine; pull on another. Code-review tooling becomes review tooling for the project's shared decisions.
• Hydration Packs: commit a .hpack file with the repo. A new developer running hydrate pack-load <file> gets the architecture, conventions, and known-broken edges in their next prompt. No onboarding read-through needed.

Recipe: plan in Opus, execute in Sonnet

1. Start in Opus for architecture, planning, or specification work.
2. When the plan is stable, run inside Claude Code: /hydrate-distill
3. Clear the current context: /clear
4. Switch to Sonnet.
5. Rehydrate the distilled session: /hydrate-last

Sonnet receives the distilled decisions and project context, without carrying the full Opus planning transcript.

FAQ