Next steps towards personalized and global knowledge graphs

[satra]

@eurunuela @MSenden @adelavega @macshine @giuliabaracc @tekrajchhetri @djarecka @puja-trivedi @neurovium @jdkent

i'm starting this thread to continue the conversation some of us started at ohbm about the evolution of the atlas and personal knowledge graphs. i'd like to think about it as a scalable, decentralized, community enterprise with an optimization function for high SNR and timely information. there are many people working on many different components of this through their own lenses, and ideally i'd like to tie things in through the lens of the brain knowledgebase. the abstract atlas was a component of this space that tied in ohbm to neuroscience abstracts, but there is much more to do, and pieces we can do relatively quickly with a convergent focus.

things tried and failed

• create a projector from minilm to the neuroscape 64d (so that i can do searches on the site against the neuroscape embeddings). don't know how close we can get the projector to be. (the learned projector was not good).

things to do

• some abstraction of claims into a knowledge model / template to the extent it's possible (really want to move from search mode to search over knowledge mode) and connections to methods and approaches many neuroscientists don't think about, and connections between problem spaces in neuroscience many computer scientists don't think about. so some abstraction over methods and problems.
• relevance vector (the current embeddings do some model of science) related to people's interests. the methods / claims embeddings still didn't do enough and i am thinking a graph embedding of some kind that matches up different dimensions. i haven't thought hard about how i would go about this yet.
• get biorxiv/psyarxiv/medrxiv in and then wire up structsense skill to extract more relevant evidence graphs.
• getting a neuroscience agent/skill in place that can leverage these graphs/data
• some formalism of evidence strength and reproducibility
• synthesis of models and frameworks or grouping of evidence for models and frameworks (mario implemented a version of this for the ohbm abstracts)

if possible i'd like to develop a scientific and engineering roadmap for this that we can assign and iterate on relatively quickly and scalably. i'm personally interested in neuroscience broadly, connections to brain health mental health, and computer science approaches that can be brought in to the space. more generally, i'm interested in developing something that is both a tool for learning and a tool for humans to ask more interesting and unanswered questions, and get clarity on how well we know something and the assumptions and provenance behind it.

feel free to ping others on this thread.

[tekrajchhetri]

@satra are we also thinking about audio2KG/voice2KG other than mining the literatures...?

@puja-trivedi We can extend the current structsense skills to also include the assertion/evidence extraction. @satra just wanted you to confirm, the priority for us is to use skills than building multi-agent systems especially where appropriate, right? If yes, we would be extending structsense skills and not update the original version structsense.

[eurunuela]

Thanks for starting this, @satra. A few thoughts from my side, coming mostly from my experience using personal knowledge management tools rather than the modeling side.

I think we should offer a way to download part of the main graph (or an embedding of it) into local systems, so that people can connect their own local graphs to the main knowledge graph. The main graph would be based on publications, abstracts and talks, while the local ones would hold all the ongoing work people are doing with their labs.

I'd also like us to offer an easy way for labs to have their own knowledge graph, where lab members can easily share their notes with each other. That's the level where a lot of the day-to-day work should happen in my opinion.

I like the idea of having skills and agents to help with this, and +1 to @tekrajchhetri on extending the structsense skills rather than building a multi-agent system.

I think a fast and reliable backend will be important, but the key thing for adoption from my ~6 years in the PKM world is that the UI is what matters most. It has to be very simple so that anyone knows how to use it, while still letting power users build their own hacks and such on top. I'd treat that as a requirement from the start, not something we polish at the end.

On process, I think we should divide and conquer and start with the backend technical features first, while keeping the UI and UX in mind. Happy to help split that into pieces we can assign.

[tekrajchhetri]

I agree with @eurunuela regarding UI. On a related note, I’d also like to share some information about MeetGraph, a tool I’ve been working on and am currently testing. It aligns well with @eurunuela's idea of sharing lab notes.

You can think of MeetGraph as something similar to Notion, but with a focus on creating and organizing knowledge graphs. The project started from the idea of generating knowledge graphs from voice-to-knowledge-graph workflows.

It provides the following features:

• meeting summary to knowledge graphs
• can use with open source models + proprietary
• team features -- one can generate team key and configure team database so every meeting summary is centralized
• if meeting generated meeting key is revoked or someone leaves the team, the meeting notes are accessible upto that particular time
• cross linking between meetings
• sharing via email, api, mcp addition to remote database relational & graph database
• linking to wikipedia, pubmed for key terms
• unless you join the team your content will be private -- you can also share it via email set your own remote database
• Note taking -- in progress

I did a test today in zoom meeting and doing more tests.

Also adding @RashiTinkX who is interning with us and is working on evidence extraction, particularly focusing on different evidentiary dimensions.

[MSenden]

@satra @eurunuela great timing on this thread. I have been working on something that speaks directly to the lab-level layer eurunuela described.

I had started building agent-maintained wikis for specific research lines / projects conducted by members of my group. It synthesizes literature across PDFs into concept pages (phenomena, models, theories brain regions, paradigms) with citation provenance and confidence levels baked in. It is designed for the kind of day-to-day work that sits below field-level infrastructure: making modeling decisions, writing methods sections, onboarding PhD students to a research line.

I am now finalizing a template NeuroWiki repo that makes it easier to construct such wikis. This template now works with Claude Code, Gemini CLI, Cursor, or any LLM agent.
It currently operates purely at the project level. The natural next layer is a lab wiki that merges project wikis. I will be working on that next and am happy to think about how it could eventually connect upward to the Brain Knowledgebase infrastructure.

I am also in the early stages of building a relatively small, local, embedding model that can provide 1024d, 64d, 3d, and 2d embeddings of scientific abstracts with the 64d embeddings exactly aligned to my neuroscape embeddings. That should remove the need for VoyageAI embeddings as a necessary step when working with new abstracts while preserving the neuroscape results.

[giuliabaracc]

Thank you @satra! I am very excited by this tool and would love to help shape it from a non-compsci/engineering point of view, more with respect to its theoretical utility and educational application. Happy to provide some input together with @macshine!

[tekrajchhetri]

I wanted to share a quick update: Meetraph is now available on GitHub:

• GitHub: https://github.com/sensein/meetgraph
• Walkthrough video: https://www.youtube.com/watch?v=GKdy1pDAbqo&feature=youtu.be

[satra]

adding @nicholsn and https://github.com/nicholsn/lokf to this thread

[satra]

and also this: https://github.com/Jaluus/ArXivAtlas

[adelavega]

There's also this from bioinformatics: https://biocypher.org/BioCypher/

[adelavega]

Kind of random, but given the relevance of this project to tech and more broad audience, would anyone be interested in creating a proposal for SXSW?

https://sxsw.com/apply/panelpicker/

Due on Sunday

[giuliabaracc]

I presented last year at SXSW Sydney and it was so much fun!! Happy to help craft something with you folks, if you are keen :) Could be something like: the need for resources that help keep pace with accelerating scientific discovery and knowledge accumulation, while making knowledge accessible to experts, novices and the public. Neuroscience could serve as the case study but this concept, I am sure, could be easily applied to other disciplines as well.

…

On 23 Jul 2026, at 3:53 AM, Alejandro de la Vega ***@***.***> wrote: adelavega left a comment (sensein/abstractatlas#66) <#66 (comment)> Kind of random, but given the relevance of this project to tech and more broad audience, would anyone be interested in creating a proposal for SXSW? https://sxsw.com/apply/panelpicker/ Due on Sunday — Reply to this email directly, view it on GitHub <#66?email_source=notifications&email_token=AHPYDX2YHBGZ6A36MFCBQQT5GD5RZA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMBUHE2TKOJWGY3KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-5049559666>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AHPYDX2OURAYRKWVWUIYVWD5GD5RZAVCNFSNUABGKJSXA33TNF2G64TZHMYTCNZXGQYTINJVHA5US43TOVSTWNBXGUZDINZRGI3TPILWAI>. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/AHPYDX2X72STOWZTVRV4TG35GD5RZA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMBUHE2TKOJWGY3KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android <https://github.com/notifications/mobile/android/AHPYDX4GLFLQYR5LVW3ON6D5GD5RZA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMBUHE2TKOJWGY3KM4TFMFZW63VHNVSW45DJN5XKKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you were mentioned.