Guide · Claude Code

Give Claude Code a memory that survives the session.

Claude Code starts every session from zero. CLAUDE.md holds standing instructions, and its auto-memory notes live on one machine, under one account. Kognite adds the missing layer: a shared, searchable long-term memory your Claude Code sessions read and write over MCP — the same memory your other tools can use too.
Start freeRead the quickstart
Why

What Claude Code forgets — and what it keeps

Claude Code has two built-in kinds of persistence, and both are local. CLAUDE.md is a static instruction file: excellent for coding standards, wrong for accumulating knowledge, because someone has to edit it. Its auto-memory directory does accumulate notes — but per machine and per account. Move to your laptop, reinstall, or share the project with a teammate's agent, and none of it follows.

A memory MCP server moves that state out of the machine. Facts, decisions, and preferences are stored once, then recalled by semantic search from any session — Claude Code on your desktop, Claude Code on the laptop, or Cursor and VS Code pointed at the same store.

Setup

Connect in one command

One command, no key to create first. The first time Claude Code reaches for a memory tool it opens a browser tab where you sign in and approve the connection, and you can revoke it later from the Kognite dashboard:

terminal
claude mcp add --transport http Kognite https://mcp.kognite.dev/mcp
  • ✓Add --scope user to make the server available in every project, not just the current one.
  • ✓Verify with claude mcp list — Kognite should show as connected.
  • ✓That's the whole setup: no npm package, no local process, no plugin. The URL is the integration.
Make it proactive

Teach Claude Code when to remember

Tools connected over MCP are available, not automatic. Two lines in CLAUDE.md turn Kognite from a tool Claude Code could call into the way it actually works:

CLAUDE.md
## Memory
Before answering anything that depends on my history, preferences, or
past decisions, call memory_search first. When I state a durable fact,
preference, or decision, persist it with memory_add.

With that in place, “we decided to use pnpm, not npm” said once in March is retrieved in July — in a different repo, on a different machine — without you re-explaining it.

What you get

More than a notes file

  • ✓Semantic recall, not grep. Hybrid vector + full-text search finds “the database decision” even when no note contains those words.
  • ✓A knowledge graph underneath. Entities and relationships with temporal edges — on every plan, including free.
  • ✓Cross-tool by construction. The same memory serves Cursor, VS Code, Zed, Windsurf, and Claude Desktop — one URL each.
  • ✓Scopes for isolation. Personal, team, and per-agent scopes keep one agent's memory from bleeding into another's.
  • ✓Reads are never metered. Quota counts memories created — searching your own memory is free on every plan.
FAQ

Common questions

Doesn't Claude Code already have memory?

It persists CLAUDE.md and local auto-memory notes — but both are per-machine and per-account. Kognite stores memory centrally, so it follows you across machines, projects, and other MCP tools.

Is memory shared between my projects?

Your choice. --scope user shares one memory across projects; the default keeps the server per-project; Kognite scopes separate personal, team, and per-agent memory on the server side.

What does it cost?

Free: 5,000 new memories/mo with unmetered search. Paid starts at $19/mo. The knowledge graph is on every plan — see pricing.

Does Kognite train models on my memories?

No — Kognite does not train models on your data. Memory formation (turning raw notes into clean, atomic facts) runs through an LLM API as part of processing; the privacy policy covers how stored data is handled.

One command. Every session remembers.

The free tier includes 5,000 new memories a month with unmetered search — enough to run Claude Code with real memory before paying anything.

Start freeSee pricing