Launched this week
Mnemcore
Turns hours of team video and notes into searchable memory
48 followers
Turns hours of team video and notes into searchable memory
48 followers
Mnemcore turns your organization’s video into searchable, shared knowledge. Capture timestamped notes, search across every recording, and ask questions to get grounded answers linked directly to the original moments. Built for teams that rely on video but need a better way to remember, find, and act on what they see.
Mnemcore
Mnemcore
Congratulations on the launch. The timestamped evidence is what makes this interesting for me. A lot of teams collect hours of video, but the useful insight still lives in someone’s notebook or memory. I could see this working well beyond sport in areas like healthcare training, manufacturing reviews, user research and field operations. One thing I’d like to understand is how Mnemcore handles disagreement. If two coaches or analysts interpret the same moment differently, does it preserve both views and show who added each observation?
Mnemcore
@os_ishmael Os, thanks for your comment.
It handles disagreement. It knows who left what note, and the interpretation of the moment is simply more evidence. Every interpretation helps our system understand what happened there. We are working closely with sports teams for this reason. Multiple experts, player-vs-coach perspectives on the same moment.
Interestingly enough we've been able to show coaches "what your players think about this", due knowing what players have said over the same moment.
Thanks for the question!
Best,
hello@mnemcore.com
the rugby-to-general-teams jump is the part I'd want to hear more about. match footage is short, physical, and the "moment" is usually obvious - a tackle, a missed pass. a lot of the other use cases you listed, like research interviews or ops reviews, are long and verbal, where the important moment is a sentence buried in an hour of talking, not a visual event. is the retrieval doing anything sport-specific under the hood right now, or is the current model already general enough that a coach's clip and a two-hour interview get indexed the same way
Mnemcore
@galdayan Gal, thanks for leaving a comment. To answer your question clearly, its domain agnostic. Any video and notes are indexed the same way, searchable and answerable. We specifically have not added any sport-specific processing anywhere in our pipeline.
Thanks for the question!
Best,
hello@mnemcore.com
@reyesjl good to know it's the same pipeline either way. that actually makes the harder case more interesting to me - for something like a two-hour interview, how does it decide what counts as the retrievable moment when there's no obvious visual cue like a tackle to anchor on? is it chunking on topic shifts in the transcript, or something else
Mnemcore
@galdayan Great question. Right now, Mnemcore does not infer the moment from a visual cue or automatically split a transcript by topic, it does not need to do this - this is like adding "tackle" tag to a clip (you cannot hope to capture everything people care about this way)...
The retrievable unit is the timestamped observation captured by the person reviewing the video. In a match, that might be “poor spacing before the turnover.” In a two-hour interview, it could be “customer explains why onboarding failed.”
We embed that observation while preserving its connection to the exact moment in the source. Later, someone can search or ask about the idea and jump directly back to where it was discussed.
That human-authored layer is intentional: we are indexing what the organization learned from the video, not merely everything that appeared or was said in it. Transcript-based ingestion and topic segmentation can become another input layer, but they are not required for the core workflow.
Your harder-case framing is exactly why I’d value your perspective. Would you be open to a quick demo? I’d love to show you the current sports workflow and discuss how it extends to interviews and operational reviews.
the grounded-answer-linked-to-the-original-moment part is what separates this from just running transcripts through a search index, most tools give you a text blob back and you still have to hunt for where in the recording it actually happened. does it handle screen shares and slides shown during the call too, or mainly spoken content?
Mnemcore
@omri_ben_shoham1 Omri, thanks for the comment!
It can handle screen shared and slides, but its not at the moment. We aren't using any VLMs yet, that's a decision we made on purpose. Its extremely cheap, by a significant amount, to work without VLM processing. That will capture slides, and screen shares, and we can upgrade to that when its needed very easily by simply adding that to our pipeline. However, we are very focused on organizational knowledge from human authored notes.
thought this is a clear next step for us, the VLMs. Hope I've answered well enough for you!
Thanks for the question!
Best,
hello@mnemcore.com
Congrats on the launch. The answer layer is the part I got curious about: when someone asks a question and gets a citation back, is retrieval scoped to the videos that person can already open? Answering across a whole workspace seems like a fun problem to get right.
Mnemcore
@vollos Hey Chalermpon,
Thanks for leaving a comment. The answer layer works both inside of a single video, and across the entire organization! You'd be surprised how far you can get with search embeddings that have high dimensional representations. They work very well, and scale even better.
We haven't used hierarchical representation of the information yet (chapters, video summaries, etc) you can further embed those too. Parent retrieval is what I learned that is called, and I tested a bit with it too, but its not what we are doing right now.
Thanks for the question!
Best,
hello@mnemcore.com
@reyesjl Parent retrieval is a good one to have tested even if you're not using it yet. The org-wide part is what I keep chewing on though: for a team that splits by role, coaches versus players say, does a query only search video the asker's already allowed into, or is there a permission check ahead of the embeddings that filters candidates before they ever reach the answer?
Mnemcore
@vollos I see what you mean. Right now, everyone within an organization can search and ask questions across the entire corpus.
We already support organization membership, roles, and invitations, but we have not yet introduced groups within an organization. Groups could allow an organization to organize members into specific teams, departments, or permission sets.
We also already have collections. Staff can add full videos to a collection, or clip specific moments, attach notes, and add those moments to a collection.
Thinking about it more, groups and collections could work together to solve this:
Groups would determine who has access, while collections would determine what content they can access.
For example, an organization could give one group access to a particular collection of videos, clips, and notes. Search and Ask would then only retrieve information from the collections that user is permitted to view.
That would let us add meaningful access control without rebuilding the entire permissions system from scratch. We could build on the organization membership, roles, invitations, and collections that already exist, then introduce groups as the layer connecting specific users to specific collections.
I had considered permissions more broadly, but I had not fully connected groups and collections as the combined solution until thinking through your question. It is especially important for larger organizations, where some information may need to be protected or hidden from particular teams.
@reyesjl Groups plus collections doing double duty, who gets in and what they can see, is a solid way to add that without ripping out what already exists. Good thread, appreciate you walking through the reasoning.
@Jose that makes sense as a sequencing call, most teams get more value from getting the text-based knowledge solid first anyway. when you do add VLM support for slides and screen shares, will that run automatically on everything going forward, or will it be something people have to opt into per-recording given it'll presumably cost more per minute processed?