Memexa · 镜我

English · 中文

A self-hosted personal memory graph over Chinese-native data — WeChat / QQ / email / documents / audio. This repository is the open demo. The full engine is a separate proprietary product.

What this is

memexa turns scattered, multi-party Chinese data into a queryable memory graph: every message is stored verbatim, extracted into structured cards, and every answer is cited back to the original sentence. It is fully self-hosted — your data never leaves your machine.

This repository ships the open demo: a small synthetic dataset and a stub extractor, so you can see the shape of the project in thirty seconds — no backend, no API key, no configuration.

Try the demo

pip install memexa
memexa demo

Six synthetic sources (WeChat / QQ / email / browser / AI chat / audio) are ingested with the stub extractor, then five sample queries run against the resulting cards — entirely in memory. This is the honest first look at what the project does.

The full engine

The demo runs a stub on synthetic data. The full memexa engine is a proprietary product and is not included in this repository. It adds:

• Live ingestion of your own data across multiple sources, incremental.
• A two-LLM extraction pipeline producing cards with per-claim citations and cross-alias canonical identities.
• A multi-channel recall stack with cross-encoder re-ranking — built for high-accuracy retrieval over messy, multi-party Chinese chat, not a single-vector lookup.
• An MCP server + CLI, so any coding agent (Claude Code, Cursor, Cline, Codex) can use your memory as a first-class tool.
• A local desktop app — runs the whole stack on your own machine.

For access to the full engine, please reach out via the repository owner's profile.

License

The demo in this repository is licensed under Apache-2.0 (see LICENSE). The full memexa engine is a separate proprietary product and is not covered by that license.