Quickstart

Two ways in. The Claude Code plugin gives you the full loop — automatic capture while you work, recall injected into every session. Any other MCP client (Claude Desktop, claude.ai, Cowork, ChatGPT, Codex CLI, Gemini CLI) connects to the same pool with recall plus on-demand remembering.

1. Create an account

Sign up at anamnesis.smtry.ai. You get a handle (your memory pool's name) and an API key — the key also derives your encryption key, so treat it like a password.

2a. Claude Code (full loop: automatic capture + recall)

/plugin marketplace add https://github.com/israelashley/anamnesis-claude-code
/plugin install anamnesis@smtry

Then run anamnesis-config once and paste in your API key and handle.

That's it. The plugin's lifecycle hooks capture your sessions deterministically — no reliance on the model deciding to call a tool — and every new session starts with your relevant memory already in context, stamped with the current date and time.

2b. Any MCP client (recall + on-demand remember)

Point the client at your personal connector URL:

https://anamnesis.smtry.ai/mcp/<your-handle>

authenticated with your API key (or OAuth where the client supports it). The connector exposes memory tools — retrieve, remember, stats, crystallize — so the assistant can recall your saved memory in any chat and save new memories when you ask it to.

3. Try it

In a fresh session, ask something your past sessions would know:

"What do we know about the payment service retry policy?"

The assistant answers from memory — with provenance, and current as of the last decision (a superseded value won't be served as current; see belief revision).

4. See your memory

Browse, search (semantic + date-aware), and delete anything at anamnesis.smtry.ai/memory. The browse view renders from an encrypted index — bodies decrypt only when you open them.

What's happening underneath

Captured sessions flow through a consolidation pipeline: raw episodes are screened (injection attempts and non-memories are dropped, not stored), distilled into salience gists, and periodically crystallized into durable long-term memory with provenance trust tiers. The mechanism vocabulary and guarantees are on the security architecture page; the measured behavior is the benchmarks.

Sharing and guests

A pool owner can issue read-only scoped tokens (memory.read) so another client — a teammate's tool, a local model — can draw on the pool without being able to write to it. Enforced at the API layer on every route; see the Trust Center.

Questions

hello@jiashley.com — private beta, and we actually read it.