Research Flywheel

Research Flywheel

Research Flywheel 是一个 local-first 的个人科研知识库飞轮。它面向长期研究者，把本地论文、PPT、笔记、每日新文献、聊天反馈和多模型审议持续炼化成可查询、可反馈、可晋升的长期科研记忆。

它不是代码 agent、训练日志分析器，也不是自动写论文机器。默认目标是先培养一个懂你领域、能引用证据、能接受反馈、能逐步形成判断力的个人科研知识系统。

论文/PPT/笔记/每日新文献
  -> paper cards / claim candidates
  -> claim ledger / self-test
  -> evidence-first 知识库对话
  -> Codex / Claude / Gemini 独立思考
  -> discuss + verdict
  -> 用户反馈
  -> reflect 晋升/降级
  -> semantic / procedural / expert profile memory

Quick Start

For AI agents opening this repository, start with [AGENTS.md](AGENTS.md). The detailed shared protocol is in [docs/agent-operating-protocol.md](docs/agent-operating-protocol.md). [CLAUDE.md](CLAUDE.md) and [GEMINI.md](GEMINI.md) are role overlays, not separate sources of truth; they should be read after the shared context. For a full external review brief, read [docs/claude-review-brief.md](docs/claude-review-brief.md). Repository changes should follow the change-impact protocol so code, docs, tests, and examples stay synchronized.

python3 -m pip install -e ".[pdf,yaml]"
rf init ./research-workspace
rf doctor --workspace ./research-workspace
rf ingest papers ./research-workspace/inbox/papers --workspace ./research-workspace
rf discover papers --candidate-pool 100 --limit 10 --workspace ./research-workspace
rf memory rebuild-index --workspace ./research-workspace
rf memory search "orthophoto building segmentation" --workspace ./research-workspace
rf claim add "Multimodal fusion may improve building separation under cross-scene shift." --evidence wiki/papers/example.md --confidence hypothesis --tags multimodal,segmentation --workspace ./research-workspace
rf claim list --status hypothesis --workspace ./research-workspace
rf daily --workspace ./research-workspace
rf ask "哪些论文思想值得迁移到我的 building segmentation pipeline?" --workspace ./research-workspace
rf council ask "这个方向是否值得深读？" --workspace ./research-workspace
rf self-test generate --topic building-instance-segmentation --workspace ./research-workspace
rf self-test answer --latest --workspace ./research-workspace
rf self-test grade --latest --workspace ./research-workspace
rf feedback add "这个 idea 值得继续观察" --target idea --workspace ./research-workspace
rf reflect --workspace ./research-workspace
rf memory inspect-claim-graph --workspace ./research-workspace
rf sync notion --workspace ./research-workspace

如果不想安装，也可以直接运行：

python3 -m rf.cli init ./research-workspace

Core Ideas

• Markdown 是权威知识底座，SQLite 保存元数据、反馈和问题历史。
• rf memory rebuild-index/search 使用 SQLite FTS5/BM25 做本地全文检索；索引是派生层，不是权威记忆。
• rf discover papers 是每日进料口：先检索并评分约 100 篇候选，再按 8 维相关度筛出 Top 10，分类保存 PDF/metadata 并入库。
• claim ledger 是判断层：每条重要结论必须有 evidence、confidence、status 和 user verdict。
• self-test 是专家考核层：定期暴露薄弱点，写入 memory/expert_profile/weak_spots.md。
• Obsidian 可以直接打开 wiki/ 和 memory/。
• Notion 是可选 dashboard，不是唯一真相来源。
• rf ask 先构造 evidence pack，再回答问题。
• rf council ask 让 Codex、Claude、Gemini 独立思考后再综合裁决。
• rf reflect 只晋升 supported claims、用户认可反馈和重复出现的问题；被否定内容会进入 expert profile。
• Hermes Agent、MCP、模型 API 都是可选外壳；Research Flywheel 保持独立。

Workspace Layout

rf init 会生成：

research-workspace/
  research-flywheel.yaml
  RESEARCH_BRIEF.md
  inbox/
    papers/
    notes/
    presentations/
  raw/
  wiki/
    papers/
    claims/
    concepts/
    questions/
    reading/
    ideas/
    writing/
    council/
    daily/
  memory/
    working/
    episodic/
    semantic/
    procedural/
    feedback/
    expert_profile/
  exports/

详见 [docs/architecture.md](docs/architecture.md) 和 [docs/configuration.md](docs/configuration.md)。

MVP Status

当前实现覆盖：

• rf init
• rf doctor
• rf ingest papers
• rf discover papers
• rf claim add/list/review
• rf daily
• rf ask
• rf council ask
• rf self-test generate/answer/grade
• rf feedback add
• rf reflect
• rf memory rebuild-index/search/inspect-claim-graph
• rf sync notion
• rf mcp serve

Graphiti 关系图适配器、真实 Notion 写入、向量索引和 Hermes skill 集成保留为可扩展接口。MVP 默认离线可用。

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: lidapengpeng
• Source: lidapengpeng/research-flywheel
• License: MIT

Install and usage instructions live in the source repository linked above.