MemStrata

A local-first, verification-first context engine for AI workflows.

MemStrata sits between you and every AI tool you use. It captures conversations from the browser, indexes your codebase locally, and serves a context block to whichever LLM you're talking to next. Your code stays on your machine. Your conversation history stays on your machine. No telemetry leaves the box.

This repository is the open-source core: the local daemon, the chat-capture browser extension, the MCP server, and the dashboard. It is MIT-licensed and fully usable on its own.

The commercial Pro tier (token-budgeted context injection through a proxy harness, money-back-guaranteed savings, IDE integration) lives in a separate, private repository and consumes this package as a PyPI dependency. See memstrata.dev if you want the paid product.

What's in this repo

Quickstart

pip install memstrata
memstrata init     # one-time interactive setup
memstrata api      # start the daemon on 127.0.0.1:8000

The daemon binds to 127.0.0.1:8000. Open http://127.0.0.1:8000/dashboard to see what's captured.

Install the browser extension from your browser's add-on store (see the browser-extension/ directory for build instructions if you want to load it unpacked).

Architecture commitments

These aren't aspirational — they're enforced in the code and tested for in CI:

1. Localhost-only binding. Every HTTP server in this repo hard-codes host="127.0.0.1". No 0.0.0.0, no LAN exposure, no remote access.

2. No TLS interception. The MCP server and the dashboard speak plain HTTP on loopback. The browser extension talks directly to provider APIs and to this daemon's loopback endpoint. There is no MITM proxy in the open-source stack.

3. Local storage only. All telemetry, all chat history, all vectors, all API keys live in ~/.memstrata/ (or $ML_DATA_DIR if set). Nothing is uploaded to a MemStrata-owned cloud service. There is no such service.

4. Telemetry never includes user content. The dashboard and the MCP server expose your data back to you. Nothing is sent off-machine.

5. The Pro tier is structurally separate. Pro code lives in a different repository under a different license. This repo has zero import statements that touch Pro code.

Provider pricing (for the dashboard's savings calculator)

The dashboard's session-level savings columns compute against a price table. By default the daemon syncs the table once per day from OpenRouter. When the network is down or OpenRouter is unreachable, the bundled memstrata/layer3/pricing/pricing_matrix.json provides a static fallback. The fallback covers the most common Claude, OpenAI, Gemini, DeepSeek, xAI, and Mistral models. Prices in the fallback are USD-per-million-tokens, last verified mid-2026.

Contributing

See CONTRIBUTING.md. The short version: open issues first for non-trivial changes, run pytest before submitting, and keep the architectural commitments above intact.

Security

See SECURITY.md. Vulnerability reports go to security@memstrata.dev, not GitHub issues.

License

MIT. See LICENSE for the full text.

Related

• Commercial Pro tier: memstrata.dev — the token-budgeting interception harness, the money-back guarantee, and the IDE extension. Proprietary, paid, separate codebase.
• Browser extension store listings: shipped from the same source tree in this repo; see browser-extension/README.md for build + sideload instructions.