What is Klara
Klara is a memory layer for Claude. Connect it once, and every session has access to everything you've chosen to save — ideas, questions, decisions, open problems, direction shifts, notes. It's always there.
Sign up, pick a plan, and you'll get your SKILL.md and the MCP link. Drop the SKILL.md to your Claude Project and connect Klara with the link. No code, no IDs, no configuration needed. Klara sets up your memory automatically.
From that point, Klara is live in your sessions. When something is worth keeping — say it or ask Claude to write it to Klara. Claude formats it, shows you what it's about to write, and you confirm. That's the memory note. No interpretation, no ambient inference. You saw it, you approved it, it's in memory.
On recall, ask Claude to pull something specific and it queries Klara directly. What comes back is exactly what was written.
Founder mode is structured: every entry captures what was decided, what was considered, what was rejected, and why — or as much of that as applies. Decisions are attributed. Open problems stay open until resolved.
Generalist mode is more free: summary and body, no required fields. You pick the mode on onboarding. When in doubt, take generalist mode.
Structured entries. What was decided, what was considered, what was rejected, and why. Decisions are attributed. Open problems stay open until resolved. Recommended for founders building with Claude.
Summary and body, no required fields. Captures anything worth saving — ideas, notes, references, observations. When in doubt, start here.
You open your Claude Project. Claude reads the SKILL.md, calls memory_read, and retrieves your last entries from Klara. Claude orients silently and opens with a one-line status: how many entries are loaded, when the last one was written, and how many open problems you have. You pick up exactly where you left off without re-explaining anything.
You've just made a decision. You say "save that." Claude identifies the entry type, formats it — summary, body, what was considered, what was rejected, why this won — and prints it to chat before writing anything. You read it, edit one line, and confirm. Claude calls memory_write with your version. Klara writes the entry and confirms back.
Three sessions later, you ask "what did we decide about the database?" Claude calls memory_read with "database" as the query. Klara returns your matching entries. Claude surfaces the relevant decision — including what was rejected and why — without you having to re-explain context or search through old chats.
Klara never sees nor stores your conversations. It only receives what you explicitly write via a tool call. No ambient writes, no interpretation.
If Klara is not available, before reconnecting it in Settings, you can try asking Claude to call her again, or refresh your chat window.
The SKILL.md file Klara generates tells Claude how to read and write your memory. Code phrases are the only safe section to personalise. Everything else should be left as generated. If memory type or attribution tag instructions are removed or changed, writes will fail silently: Claude will appear to save something, but nothing reaches Klara. If memory stops working after an edit, re-download the original SKILL.md from klara.tools/settings.