Live Demo — fully client-side

Semantic memory, running entirely in your browser

A real vector database (sqlite-vec compiled to WebAssembly) and a real embedding model (all-MiniLM-L6-v2, ONNX) run on this page. Store memories, search them by meaning — nothing is uploaded, ever.

Search enables once the embedding model (~23 MB) finishes loading. It is fetched from this site only — no CDN, no third parties.

Stored memories

    Verify it yourself: open your browser's network tab. After the initial same-origin asset load, this page makes no further requests — embedding, storage and search all happen locally. Reloading the page discards everything.

    This is the real stack, shrunk to a page

    The demo uses the same components as the self-hosted MCP Memory Service — just without persistence, consolidation, or multi-client access.

    In this demosqlite-vec (WASM, in-memory)
    →
    In the servicesqlite-vec backend with persistent storage, WAL, hybrid cloud sync
    In this demoall-MiniLM-L6-v2 ONNX in the browser
    →
    In the servicethe same ONNX embedding model, server-side — no GPU, no API keys
    In this demoone page, one session
    →
    In the serviceMCP + REST + OAuth 2.1 for 25+ AI clients, knowledge graph, consolidation

    Or hand the trial to your coding agent

    Paste this into Claude Code in any directory. It installs the service into one new folder, proves that storing and searching work, and registers it for that folder only. Read it first: it is everything your agent will be told.

    Set up a trial of MCP Memory Service (https://github.com/doobidoo/mcp-memory-service) for Claude Code.
    
    Keep everything inside a new folder named mcp-memory-trial in the current directory. Tell me what you are about to do and wait for my OK before you install or register anything. Do not change any other configuration, and leave an existing MCP server named "memory" alone.
    
    1. Find Python 3.11 or newer. If there is none, stop and tell me. Do not install Python.
    2. In mcp-memory-trial, create a virtual environment named .venv and install into it:
       pip install "mcp-memory-service[sqlite]"
    3. Run the health check with the database inside the folder. TRIAL stands for the absolute path of mcp-memory-trial:
       MCP_MEMORY_STORAGE_BACKEND=sqlite_vec MCP_MEMORY_BASE_DIR=TRIAL/data MCP_MEMORY_SQLITE_PATH=TRIAL/data/sqlite_vec.db TRIAL/.venv/bin/memory status --deep
       The first run downloads the embedding model, about 170 MB, to ~/.cache/mcp_memory. Show me the output. It has to contain "Backend class: ONNXEmbeddingModel" and "Service is healthy". If it does not, stop and show me what it says.
    4. From inside mcp-memory-trial, register the server for this folder only:
       claude mcp add --scope project memory-trial -e MCP_MEMORY_STORAGE_BACKEND=sqlite_vec -e MCP_MEMORY_BASE_DIR=TRIAL/data -e MCP_MEMORY_SQLITE_PATH=TRIAL/data/sqlite_vec.db -- TRIAL/.venv/bin/memory server
    5. List what you created. Then tell me to start a new Claude Code session inside mcp-memory-trial, approve the memory-trial server, and try these two messages:
       Use memory-trial to remember that our staging database is Postgres 16 behind PgBouncer.
       Use memory-trial: what do you know about our staging setup?
    6. Tell me how to undo it: delete the mcp-memory-trial folder. Say whether ~/.cache/mcp_memory existed before this trial; only if it did not, I can delete it as well.

    “Open in Claude Code” hands this exact text to the Claude desktop app, where you review it before sending; the app has to be installed. ChatGPT has no documented way to receive a prompt through a link, so that button copies the text and opens ChatGPT for you to paste. A chat window cannot install anything: it can walk you through the steps, while a coding agent carries them out.

    Tested with the Claude Code command line on macOS; the hand-over to the desktop app passes the same text but was not run there. The folder takes about 250 MB. A first install also downloads the 170 MB embedding model to ~/.cache/mcp_memory, which is outside the folder.