Connectore
Connectore is a local-first second brain. You write short notes (lecture notes, textbook concepts, news snippets), and Gemma 4, running fully on a laptop through Ollama, links them across subjects, tells the story behind them, and then checks its own story against your notes for facts it dropped.
The problem: Notes live in silos. Negative feedback in Control Systems, TCP congestion control in Computer Networks, and the RBI raising the repo rate to curb inflation are all the same idea (a feedback loop), but they share almost no words, so similarity-based tools miss the link. On top of that, AI summaries quietly drop caveats: a fluent summary isn’t necessarily a faithful one.
What it does:
-
Connect: Each note costs two Gemma calls. Gemma tags it with entities and abstract patterns (e.g. “feedback loop”, “oscillation”). A SQL query shortlists up to 8 older notes, then one batched Gemma call judges each candidate and returns a relation type (same_idea, causes, consequence_of, continuation, contradicts), a strength from 1 to 5, and a reason citing a fact from both notes. Only links with strength 4 or more that point to real shortlisted notes are kept, so the model can’t invent connections.
-
Graph + stories: An Obsidian-style force graph with links coloured by type; hover a link to read Gemma’s reason. Click a note to light up its thread and stream a short story that ties the notes together.
-
Drag to connect: Drop one note onto another and Gemma judges the pair. It either draws the link or explains what separates the two notes.
-
Verify: Every story gets a faithfulness report:
- a deterministic regex check for numbers, units, dates, negations, “unless”/“only if” conditions and bounds
- a QAGS-style fact quiz: Gemma extracts must-keep facts with exact quotes, then answers each one using only the story
Each fact is marked Appears preserved, Needs review or Possible mismatch. There’s no fake “verified” badge.
Measured on our laptop (gemma4:e4b, Ollama 0.40.1):
- ~48–50 tok/s
- 3–6 s to add and link a note
- the unrelated “photosynthesis” note correctly gets 0 links
- on a deliberately sabotaged story, the verifier flagged all four dropped cues (“at most”, “unless”, “₹2”, “14-day”)
Everything runs locally: no cloud API, and your notes never leave your machine. MIT licensed.
Built by Team WInDaChat: V Aravindhan, Siva Krithick, Sahesh Karthikeyan, Sathyanarayanan.
Built with
AI tools
Prize Categories Entered
- Submitted
Best Use of Gemma 4
- Submitted

Best Open-Source AI Project