Memexa · 镜我

English · 中文

Memory layer for AI agents and humans, on Chinese-native data. Self-hosted memory graph over WeChat / QQ / 飞书 / 钉钉 group chats, Chinese email, and Chinese audio. Verbatim storage plus structured extraction; queries return cards with per-claim citations back to the original sentence.

🤖 AI-agent compatible by design. Most real usage is an AI agent (Claude Code, Cursor, Cline, or one you wrote yourself) invoking memexa as a subprocess to answer questions on the user's behalf. The fourteen query subcommands are a small protocol; the contract agents follow is in docs/for_agents.md. Native MCP integration arrives in v0.5; the current first-class path is shell subprocess with --json output.

Quickstart

Two starting points; pick whichever describes you.

Humans — 30-second visual

pip install memexa
memexa demo

You will see a synthetic conversation set ingested from six sources with the stub extractor, followed by five example queries printed to your terminal — quick, arc, timeline, pending, topic. No backend, no LLM, no configuration. This is the honest first look at what the project does.

AI agents — subprocess CLI today, MCP in v0.5

# Agents already work today via subprocess:
pip install memexa
memexa quick "<your question>" --json   # structured output for agent parsing
memexa arc "<person>" --json
# ... fourteen subcommands total, all with --json mode (v0.1.x)

The fourteen subcommands plus seven hard rules in docs/for_agents.md are the agent contract. Native MCP integration (memexa-mcp server + .mcp.json snippet) arrives in v0.5; until then shell subprocess is the first-class path and it works in any agent that has a shell tool.

What's next for both

To ingest your own data, configure an LLM provider and pick one source. docs/quickstart.md walks through Tier 1 (5 minutes, one source) and Tier 2 (30 minutes, full production deployment with cron + dashboard).

What you can ask

Question pattern	Subcommand	Returns
Who is Alice?	arc "Alice"	Relationship arc, 8 fan-out variants across sources
What was the whole story behind X?	topic "Mac purchase"	80–200 cards with citations
What did Y professor want?	person "Y professor"	Profile article + recent events
What is project X across all sources?	project "X"	Cross-source pulse, 4 source groups
What is on my plate?	pending	Active commitments from calendar
What did this period look like?	timeline --start ... --end ...	Chronological card list
Synthesise an answer	reflect "question"	LLM-synthesised Markdown

Fourteen subcommands total. Decision table and composition patterns are in docs/usage_guide.md. See also docs/5_phase_query.md for the state- inference workflow used on yes/no questions.

Why memexa instead of OpenHuman / MemPalace / ReMe?

In short: verbatim raw storage + LLM-extracted V2 envelope + per-claim evidence_quotes citation + cross-alias canonical id, all on Chinese-IM-native data sources the adjacent projects do not target.

The full per-capability comparison and the five user scenarios memexa serves live in docs/why.md.

Architecture, one screen

   WeChat ─┐                                              ┌─► "Who is X?"           (arc + quick)
   QQ     ─┤                                              ├─► "Group activity last week?" (topic + trends)
   Email  ─┼──► two-LLM extract ──► PG + pgvector ──┤
   Browser─┤    (gate+extract)      memory graph        ├─► "Project X status?"    (project + timeline)
   AI chat─┤                                              ├─► "What does Y want?"   (person)
   Audio  ─┘                                              └─► "My pending actions?" (pending)
        ↑                                                       ↑
   your raw data                                          14 query subcommands
   (local, fully self-hosted)                             (cross-source composable)

Full architecture in docs/architecture.md.

Documentation

Topic	Link
Quickstart (3-tier path: 30 s → 5 min → 30 min)	docs/quickstart.md
Architecture	docs/architecture.md
Why memexa (vs OpenHuman / MemPalace; 5 user scenarios)	docs/why.md
Cost estimation (DeepSeek / GPT-4o / Claude monthly)	docs/cost.md
14 query subcommands in depth	docs/usage_guide.md
5-phase state inference	docs/5_phase_query.md
Full environment variables	docs/configuration.md
FAQ / troubleshooting	docs/faq.md · docs/troubleshooting.md
Per-source onboarding	docs/integrations/
Cross-platform deployment	docs/deployment/
Example walkthroughs (synthetic data)	examples/demo_dataset/walkthroughs/
Case studies	docs/case_studies/
For AI agents (MCP / integration spec)	docs/for_agents.md
Roadmap	ROADMAP.md
Contributing	CONTRIBUTING.md
Security policy	SECURITY.md
Governance	GOVERNANCE.md

Two ways to run the LLM

memexa's core is a two-LLM gate-extract pipeline. The OSS ships everything you need to run it locally with any OpenAI-compatible endpoint.

# Default: bundled prompts + your own LLM provider
export MEMEXA_EXTRACTOR_TIER=bundled

# BYO: bring your own prompt for advanced tuning
export MEMEXA_EXTRACTOR_TIER=byo
export MEMEXA_PROMPT_PATH=/path/to/your_prompts.py

Recommended provider for Chinese workloads is DeepSeek V4 Flash (gate)

• V4 Pro (extractor) — typical cost is ¥0.30 per 1 000 messages. GPT-4o and Claude 4.x are supported but cost 5–10× more. See docs/cost.md for the full breakdown.

License

Apache 2.0. See LICENSE. OSS core stays Apache 2.0 forever.