Using Sci

Once installed, most of what Sci does is invisible — the proxy runs in the background and memory flows automatically. You keep using Claude Code exactly as before. Below is what’s happening underneath, and the few things worth doing deliberately.

Your first session, end to end

1. Open a new terminal (so the proxy environment is inherited) and run claude in any project.
2. Work normally. Behind the scenes, every request is routed through Sci, anonymized, and the reply de-anonymized — you won’t notice a thing except, on the very first request, a brief pause while the embedding model downloads.
3. Tell Claude something worth keeping: “Remember we’re standardizing on Vitest for this repo.”
4. Tomorrow, open a session in a different directory and ask: “What testing framework are we using?” The answer is already there. That’s memory injection — the moment it becomes visible.

What happens automatically

Every time you run claude in a terminal:

1. Proxy intercepts the session. Claude Code’s HTTPS_PROXY is set to http://127.0.0.1:3001, so all requests route through Sci before reaching Anthropic.
2. PII is anonymized outbound and restored inbound — you never see the substitution.
3. Memory tools are available to Claude over MCP: memory_recall, memory_store, memory_identity, memory_status.

You don’t need to do anything for this to work.

Confirming it’s active

sci status     # ok: true, plus memory counts
sci verify     # live before/after privacy check

Asking Claude to recall context

Claude won’t call memory tools unless instructed. At the start of a session, ask:

“Check what you know about me and my active projects.”

For automatic recall on every session, add this to ~/.claude/CLAUDE.md:

## Memory
At the start of each session, call `memory_recall` with a query relevant
to the current task, and `memory_identity` with no arguments.

Storing memories

Ask Claude directly:

“Remember that we chose SQLite over Postgres for the alpha.”

Claude calls memory_store; the memory is embedded locally and persisted.

Seeding from past conversations

# Export from claude.ai → Settings → Privacy → Export Data
sci import --claude ~/Downloads/conversations.json

The four MCP tools

See What to test for the specific things we’d love feedback on.