Skip to content
Claspt Docs

AI tools: memory

Claspt is an MCP server. Any MCP client can keep its memory in your vault, in markdown pages you can read and edit, and, with your approval, use the credentials in it.

Connect a tool

Section titled “Connect a tool”
Terminal window
claspt mcp install claude-code claude-desktop

One command writes Claspt into each named client’s config with one shared key for every tool on this machine. Name any of claude-code, claude-desktop, cursor, windsurf, gemini-cli, codex. Then restart the tool.

  • --secrets issues a key that can read secret values (with your approval; see AI tools: credentials). Without it the key is Notes scope: pages, memory and search, with secret values redacted.
  • --separate gives the named tools a key of their own, so one can be revoked alone.
  • --project writes the project-scoped config in the current directory (Claude Code and Cursor) with a key limited to that project’s memory.

The key is registered in the vault the app has open, and it never appears on screen. Every key is listed under Settings › Integrations › API Clients and can be revoked there.

When something is off

Section titled “When something is off”
Terminal window
claspt mcp doctor

lists every AI tool config on the machine with the hint of the key it holds and whether the open vault and the running app accept it. A refused request in the tool says the same: which key, and the two commands that fix it.

Memory

Section titled “Memory”

Memory is filed per project. Pin a project with one committed file:

Terminal window
claspt namespace init

Every clone, machine and subfolder then agrees on which memory is which; a shared global namespace holds what applies everywhere. The tool’s first call, memory_guide, tells it the conventions and where it is.

What the tool can do with memory:

  • Write and read pages of three kinds: episodic (what happened), semantic (what is true), procedural (how things are done). A memory can say when it stopped being true and what replaced it; readers see it marked stale.
  • Append to a running log without rewriting the page, and refuse to overwrite a page another tool changed in between.
  • Read the latest part of a long page, and compact it: older sections move to a dated archive that search still finds.
  • Search across every project it may use. With Ollama running on your machine, hits are ranked by meaning; otherwise by keyword. Nothing leaves the machine, and the result says which ranking was used.
  • Import what it already has: claspt memory import reads mem0, Letta and Zep exports, and CLAUDE.md or Cursor rules files, dry run first.

You review what it wrote

Section titled “You review what it wrote”

Anything a tool writes is marked unreviewed. When any tool reads it back, the text is fenced with markers saying it is unreviewed data from another session, not instructions, with the writer’s name. Settings › Agent Memory shows every project’s memory: each page’s kind, who wrote it, how often it is read, and what is waiting for you. Mark a page reviewed with one click. Only the app can do that; no API call can.

This makes injected text visible and labelled. The model reading it still has to honour the label.

In the tool’s own menus

Section titled “In the tool’s own menus”

Memory pages are exposed as MCP resources, so clients that show resources can attach a page to a chat, and three prompts are ready: start a session, end a session, review memory.

From the command line and code

Section titled “From the command line and code”

claspt memory list|read|append|search follow the project you are in. The Python (pip install claspt) and TypeScript (npm install @claspt/sdk) SDKs carry the same six tools for LangChain, the OpenAI Agents SDK and CrewAI.