[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-80683":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":15,"subscribersCount":15,"size":15,"stars1d":14,"stars7d":13,"stars30d":13,"stars90d":15,"forks30d":15,"starsTrendScore":16,"compositeScore":17,"rankGlobal":10,"rankLanguage":10,"license":18,"archived":19,"fork":19,"defaultBranch":20,"hasWiki":21,"hasPages":19,"topics":22,"createdAt":10,"pushedAt":10,"updatedAt":33,"readmeContent":34,"aiSummary":35,"trendingCount":15,"starSnapshotCount":15,"syncStatus":14,"lastSyncTime":36,"discoverSource":37},80683,"memexa","labazhou2024\u002Fmemexa","labazhou2024","Self-hosted Chinese personal memory graph. Six sources, two LLMs, one graph.","https:\u002F\u002Fgithub.com\u002Flabazhou2024\u002Fmemex",null,"Python",50,3,2,0,6,1.81,"Apache License 2.0",false,"main",true,[23,24,25,26,27,28,29,30,31,32],"bge-m3","chinese-nlp","cli","knowledge-graph","llm-pipeline","personal-memory","pgvector","postgresql","retrieval-augmented-generation","self-hosted","2026-06-12 02:04:05","\u003C!--\nrepository-topics:\n  - personal-memory\n  - knowledge-graph\n  - chinese-nlp\n  - retrieval-augmented-generation\n  - self-hosted\n  - postgresql\n  - pgvector\n  - bge-m3\n  - llm-pipeline\n  - cli\n  - deliverable-factory\n  - action-card\n  - report-generation\n-->\n\n# Memexa · 镜我\n\n**English** · [中文](README.zh.md)\n\n> **Memory layer for AI agents and humans, on Chinese-native data.**\n> Self-hosted memory graph over WeChat \u002F QQ \u002F 飞书 \u002F 钉钉 group chats,\n> Chinese email, and Chinese audio. Verbatim storage plus structured\n> extraction; queries return cards with per-claim citations back to\n> the original sentence.\n>\n> 🤖 **AI-agent compatible by design.** Most real usage is an AI agent\n> (Claude Code, Cursor, Cline, or one you wrote yourself) invoking\n> memexa as a subprocess to answer questions on the user's behalf.\n> The fourteen query subcommands are a small protocol; the contract\n> agents follow is in [`docs\u002Ffor_agents.md`](docs\u002Ffor_agents.md).\n> Native MCP integration arrives in v0.5; the current first-class\n> path is shell subprocess with `--json` output.\n\n[![CI](https:\u002F\u002Fgithub.com\u002Flabazhou2024\u002Fmemexa\u002Factions\u002Fworkflows\u002Fci.yml\u002Fbadge.svg)](https:\u002F\u002Fgithub.com\u002Flabazhou2024\u002Fmemexa\u002Factions\u002Fworkflows\u002Fci.yml)\n[![CodeQL](https:\u002F\u002Fgithub.com\u002Flabazhou2024\u002Fmemexa\u002Factions\u002Fworkflows\u002Fcodeql.yml\u002Fbadge.svg)](https:\u002F\u002Fgithub.com\u002Flabazhou2024\u002Fmemexa\u002Factions\u002Fworkflows\u002Fcodeql.yml)\n[![License: Apache 2.0](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FLicense-Apache_2.0-blue.svg)](LICENSE)\n[![Python 3.10+](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fpython-3.10+-blue.svg)](pyproject.toml)\n[![PyPI](https:\u002F\u002Fimg.shields.io\u002Fpypi\u002Fv\u002Fmemexa?label=PyPI)](https:\u002F\u002Fpypi.org\u002Fproject\u002Fmemexa\u002F)\n[![PII scan](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPII%20residual-0%20matches-success.svg)](scripts\u002Ffull_pii_scan.sh)\n\n## Quickstart\n\nTwo starting points; pick whichever describes you.\n\n### Humans — 30-second visual\n\n```bash\npip install memexa\nmemexa demo\n```\n\nYou will see a synthetic conversation set ingested from six sources\nwith the stub extractor, followed by five example queries printed to\nyour terminal — `quick`, `arc`, `timeline`, `pending`, `topic`. No\nbackend, no LLM, no configuration. This is the honest first look at\nwhat the project does.\n\n### AI agents — subprocess CLI today, MCP in v0.5\n\n```bash\n# Agents already work today via subprocess:\npip install memexa\nmemexa quick \"\u003Cyour question>\" --json   # structured output for agent parsing\nmemexa arc \"\u003Cperson>\" --json\n# ... fourteen subcommands total, all with --json mode (v0.1.x)\n```\n\nThe fourteen subcommands plus seven hard rules in\n[`docs\u002Ffor_agents.md`](docs\u002Ffor_agents.md) are the agent contract.\nNative MCP integration (`memexa-mcp` server + `.mcp.json` snippet)\narrives in v0.5; until then shell subprocess is the first-class path\nand it works in any agent that has a shell tool.\n\n### What's next for both\n\nTo ingest your own data, configure an LLM provider and pick one\nsource. [`docs\u002Fquickstart.md`](docs\u002Fquickstart.md) walks through Tier\n1 (5 minutes, one source) and Tier 2 (30 minutes, full production\ndeployment with cron + dashboard).\n\n## What you can ask\n\n| Question pattern | Subcommand | Returns |\n|---|---|---|\n| Who is Alice? | `arc \"Alice\"` | Relationship arc, 8 fan-out variants across sources |\n| What was the whole story behind X? | `topic \"Mac purchase\"` | 80–200 cards with citations |\n| What did Y professor want? | `person \"Y professor\"` | Profile article + recent events |\n| What is project X across all sources? | `project \"X\"` | Cross-source pulse, 4 source groups |\n| What is on my plate? | `pending` | Active commitments from calendar |\n| What did this period look like? | `timeline --start ... --end ...` | Chronological card list |\n| Synthesise an answer | `reflect \"question\"` | LLM-synthesised Markdown |\n\nFourteen subcommands total. Decision table and composition patterns\nare in [`docs\u002Fusage_guide.md`](docs\u002Fusage_guide.md). See also\n[`docs\u002F5_phase_query.md`](docs\u002F5_phase_query.md) for the state-\ninference workflow used on yes\u002Fno questions.\n\n## Why memexa instead of OpenHuman \u002F MemPalace \u002F ReMe?\n\nIn short: verbatim raw storage + LLM-extracted V2 envelope + per-claim\n`evidence_quotes` citation + cross-alias canonical id, all on\nChinese-IM-native data sources the adjacent projects do not target.\n\nThe full per-capability comparison and the five user scenarios memexa\nserves live in [`docs\u002Fwhy.md`](docs\u002Fwhy.md).\n\n## Architecture, one screen\n\n```\n   WeChat ─┐                                              ┌─► \"Who is X?\"           (arc + quick)\n   QQ     ─┤                                              ├─► \"Group activity last week?\" (topic + trends)\n   Email  ─┼──► two-LLM extract ──► PG + pgvector ──┤\n   Browser─┤    (gate+extract)      memory graph        ├─► \"Project X status?\"    (project + timeline)\n   AI chat─┤                                              ├─► \"What does Y want?\"   (person)\n   Audio  ─┘                                              └─► \"My pending actions?\" (pending)\n        ↑                                                       ↑\n   your raw data                                          14 query subcommands\n   (local, fully self-hosted)                             (cross-source composable)\n```\n\nFull architecture in [`docs\u002Farchitecture.md`](docs\u002Farchitecture.md).\n\n## Documentation\n\n| Topic | Link |\n|---|---|\n| Quickstart (3-tier path: 30 s → 5 min → 30 min) | [`docs\u002Fquickstart.md`](docs\u002Fquickstart.md) |\n| Architecture | [`docs\u002Farchitecture.md`](docs\u002Farchitecture.md) |\n| Why memexa (vs OpenHuman \u002F MemPalace; 5 user scenarios) | [`docs\u002Fwhy.md`](docs\u002Fwhy.md) |\n| Cost estimation (DeepSeek \u002F GPT-4o \u002F Claude monthly) | [`docs\u002Fcost.md`](docs\u002Fcost.md) |\n| 14 query subcommands in depth | [`docs\u002Fusage_guide.md`](docs\u002Fusage_guide.md) |\n| 5-phase state inference | [`docs\u002F5_phase_query.md`](docs\u002F5_phase_query.md) |\n| Full environment variables | [`docs\u002Fconfiguration.md`](docs\u002Fconfiguration.md) |\n| FAQ \u002F troubleshooting | [`docs\u002Ffaq.md`](docs\u002Ffaq.md) · [`docs\u002Ftroubleshooting.md`](docs\u002Ftroubleshooting.md) |\n| Per-source onboarding | [`docs\u002Fintegrations\u002F`](docs\u002Fintegrations\u002F) |\n| Cross-platform deployment | [`docs\u002Fdeployment\u002F`](docs\u002Fdeployment\u002F) |\n| Example walkthroughs (synthetic data) | [`examples\u002Fdemo_dataset\u002Fwalkthroughs\u002F`](examples\u002Fdemo_dataset\u002Fwalkthroughs\u002F) |\n| Case studies | [`docs\u002Fcase_studies\u002F`](docs\u002Fcase_studies\u002F) |\n| **For AI agents (MCP \u002F integration spec)** | [`docs\u002Ffor_agents.md`](docs\u002Ffor_agents.md) |\n| Roadmap | [`ROADMAP.md`](ROADMAP.md) |\n| Contributing | [`CONTRIBUTING.md`](CONTRIBUTING.md) |\n| Security policy | [`SECURITY.md`](SECURITY.md) |\n| Governance | [`GOVERNANCE.md`](GOVERNANCE.md) |\n\n## Two ways to run the LLM\n\nmemexa's core is a two-LLM gate-extract pipeline. The OSS ships\neverything you need to run it locally with any OpenAI-compatible\nendpoint.\n\n```bash\n# Default: bundled prompts + your own LLM provider\nexport MEMEXA_EXTRACTOR_TIER=bundled\n\n# BYO: bring your own prompt for advanced tuning\nexport MEMEXA_EXTRACTOR_TIER=byo\nexport MEMEXA_PROMPT_PATH=\u002Fpath\u002Fto\u002Fyour_prompts.py\n```\n\nRecommended provider for Chinese workloads is DeepSeek V4 Flash (gate)\n+ V4 Pro (extractor) — typical cost is **¥0.30 per 1 000 messages**.\nGPT-4o and Claude 4.x are supported but cost 5–10× more.\nSee [`docs\u002Fcost.md`](docs\u002Fcost.md) for the full breakdown.\n\n## License\n\nApache 2.0. See [`LICENSE`](LICENSE). OSS core stays Apache 2.0\nforever.\n","Memexa 是一个自托管的中文个人记忆图谱项目，旨在为AI代理和用户提供基于中文原生数据的记忆层。它支持从微信、QQ、飞书、钉钉群聊以及中文邮件和音频中提取信息，并将这些信息存储为结构化数据。该项目利用了两个大型语言模型（LLM）来增强检索生成能力，同时采用PostgreSQL和pgvector技术构建知识图谱。用户可以通过命令行界面（CLI）与系统交互，执行查询操作时能够获得带有引用原文句子的详细回复卡片。Memexa特别适用于需要处理大量中文文本数据并希望从中快速准确获取信息的场景，如学术研究、企业内部知识管理等。","2026-06-11 04:01:37","CREATED_QUERY"]