visp-memory MCP server

Visp Memory

Project memory for coding assistants. Keep decisions, warnings, and useful history across sessions. Retrieve relevant context with source citations, inspect why it matched, and maintain the store from a local dashboard.

SQLite and keyword search work without a model or an external database. An optional Neo4j backend is available as an opt-in beta.

flowchart LR
    A["Project history"] --> B["Memory store
Sources and notes"]
    B --> C["Cited context"]
    C --> D["Assistant"]
    D -->|"Records outcomes"| A

Your assistant owns execution and completion decisions. Memory supplies context; it does not grant permission or certify that work is finished.

Start with the dashboard

Run the published Linux amd64 image with persistent storage:

docker run -d --name visp-memory \
  -p 127.0.0.1:8000:8000 \
  -v visp-memory-data:/data \
  -e VISP_MEMORY_EMBEDDING_PROVIDER=noop \
  ghcr.io/djkeshawa/visp-memory:0.9.0
docker logs visp-memory

Open the dashboard. On a fresh store, use the one-use setup link in the logs to create your administrator account. There is no shared default password. See installation for Compose, upgrades, Python, and standalone downloads.

Dashboard overview with sample data. The dashboard starts in dark mode; use the sidebar theme control to switch appearance.

Use it from your project

Requires Python 3.10+. Run these commands in your repository:

pip install "visp-memory[mcp,capture]"
visp-memory init
visp-memory decision "Use SQLite locally" "No database service needed"
visp-memory recall "local database"
visp-memory brief "implement local storage"

init creates project configuration and imports selected Git history. Connect your assistant separately using the MCP and hooks guide. Use visp-memory doctor to inspect configuration and integration problems. To try capture and recall without touching your project, run visp-memory demo.

What you can do

• Recall project knowledge: search decisions and history, or prepare a cited task brief.
• Understand results: inspect matching words, retrieval method, sources, and quality flags.
• Maintain memories: preview dreaming cycles, schedule exact-duplicate cleanup, and review related or expired notes. Applied changes can be undone.
• Follow task progress: receive explicit completion reports from your assistant or workflow, with evidence and history.
• Keep control of data: use local storage, scoped access, and backup and restore.

The supported path is one developer working in one local repository. Dashboard/API and some integrations are still early; team and graph features are frozen. Check feature status for the limits of each surface.

Documentation

Start with the documentation index, or go directly to:

Learn	Configure	Integrate
Architecture and data flow	Installation	MCP and hooks
Trust and privacy	Accounts and tokens	Workflow reports
Dreaming	Storage and embeddings	Machine contracts

Evidence and limits

Benchmarks describe deterministic selection and poisoning fixtures, including missed memories and other negative results. They do not establish better coding outcomes or superiority over another memory product. Published selection figures are checked against measured output by tests/docs/test_published_figures.py.

Contribute and license

See Contributing for development and submission requirements. Licensed under Apache 2.0; contributions require a DCO sign-off and the Contributor License Agreement.