UReKA: Your Personal Research Knowledge Management Agent

Ingests annotated research sources (Zotero PDFs, Notion notes, Obsidian notes, arXiv papers, web pages), extracts structured knowledge, and builds an interlinked knowledge base of paper pages and concept pages — queryable in natural language. Also supports building full learning courses with /goal → /curriculum → /tutor.

---

Architecture

sources (Zotero PDFs, Notion, Obsidian, arXiv, web)
    │
    ▼  ingest — /zotero, /notion, /obsidian, /alphaxiv, /web
    ├──▶ sources/   objective text
    └──▶ notes/     personal notes + pdf annotations
              │
              ▼  /collate <topic|paper> — synthesise sources + notes into a page
          papers/<slug>.md   or   concepts/<slug>.md
          (objective summary + personal notes/annotations woven in)
              │
   /edit ─────┤  co-edit any papers/ or concepts/ page in place
              │
              ▼  /ask — synthesise a cited answer from the knowledge base (+ flag gaps)

course/<slug>/   — a learning course, three-stage pipeline:
   /goal <topic>       → goal.md   (scope + knowledge audit)
   /curriculum <slug>  → plan, schedule, modules, per-course library
   /tutor <slug>       → session-by-session teaching, flashcards, mastery tracking

New here? See the demo branch for example outputs — sources/notes ingested and collated into papers/concepts.

Setup

Python (retrieval index — no API key needed)

python -m venv .venv && .venv/bin/pip install -r .claude/requirements.txt

Skills call Python via sh .claude/src/pyrun …, which always uses .venv/bin/python — no activation needed.

Environment variables

cp .env.example .env

Then edit .env — at minimum set your contact email for polite API access:

PRIOR_LLM_BACKEND=claude-code   # uses your local Claude Code login, no API key needed
PRIOR_CONTACT_EMAIL=you@uni.com
PRIOR_DATA_DIR=.cache

See .env.example for all options. .env is gitignored — never commit it.

MCP servers (for Zotero + Notion ingestion)

The project ships an .mcp.json that registers three MCP servers. Zotero and pdf-tools are picked up automatically — approve them when Claude Code prompts on first session.

Notion uses HTTP + OAuth and must be added manually once:

claude mcp add --transport http notion https://mcp.notion.com/mcp

Then run /mcp in Claude Code to authenticate. After that it's persisted and you don't need to run it again.

• zotero — local Zotero library access for /zotero (needs the Zotero desktop app running; zotero-mcp-server is installed by requirements.txt).
• pdf-tools — local PDF body-text + ink-annotation extraction for /zotero (pdf_tools.py, also installed via requirements.txt).
• notion — live Notion workspace fetch for /notion (HTTP MCP, registered above).

Day-to-day usage

All work is done through Claude Code skills:

# Ingest a source
/notion "My reading note title"         live fetch from Notion workspace
/obsidian /path/to/your/note.md         any local .md file
/zotero                                  pick a PDF from Zotero
/alphaxiv 2507.05024                     arXiv paper via AlphaXiv API
/web https://en.wikipedia.org/wiki/...   vetted web page with credibility score

# Build knowledge pages
/collate "attention is all you need"     synthesise sources + notes → papers/ or concepts/
/edit papers/attention.md               co-edit a page in place

# Query the base
/ask "how does ReAct reduce hallucination?"
/retrieve RLHF                           list relevant notes (no synthesis)

# Build a learning course
/goal "understand diffusion models" by 2026-08-01 5 h/week
/tutor diffusion-models

Structure

Path	What
sources/	Ingested objective text (zotero_*, alphaxiv_*, web_*)
notes/	Ingested personal notes + annotations (zotero_*, notion_*, obsidian_*)
papers/	Paper pages — objective summary + personal notes
concepts/	Concept pages synthesised across the base
course/<slug>/	Learning course: goal.md, plan.md, schedule.md, modules/, progress.md, library/
explore_library/	Standalone /autoexplore corpus (separately indexed, not the personal base)
.claude/src/	All Python tooling — BM25 retrieval, ingestion helpers, course utilities, web resources (pure Python, no LLM)
.claude/skills/	Claude Code skills (/collate, /edit, /ask, /retrieve, /goal, /curriculum, /tutor, ingestion helpers)
.claude/AGENTS.md	Full architecture and conventions

Learn More

• BLOG.md — overview of the problem, what we built, and design decisions
• ROADMAP.md — current state, next steps, and how to contribute

Acknowledgements

Built during Agents4Academia, 15–26 June 2026.