MindMeld

» see mindmeld.com as it lived, on the wayback machine

~ the obituary ~

The ending is four commits. On 28 July 2026, in the fourteen minutes between 18:33 and 18:47 GMT, somebody at Cisco pushed four changes to the cisco/mindmeld repository. Three of them rewrote the README. The fourth was titled fix year in README, because whoever was writing the death notice had put the wrong date in it and came back to correct the record on a project nobody was going to read again. The repository was archived that evening.

The notice they left sits above everything the project had ever documented about itself, under the heading READ THIS FIRST. "The MindMeld SDK is no longer maintained. Though once a state-of-the-art conversational AI platform, MindMeld has long since been surpassed by technologies enabled by LLMs." It advises anyone still using the library to evaluate alternatives, notes that there had been no release since November of 2022, and confirms there will not be another one. It closes: "Thank you to all our users and contributors."

It started as Expect Labs, founded in 2011 by Tim Tuttle, who had come out of MIT. The first product was an iPad app that listened to the conversation in the room and put relevant information on screen before anyone thought to search for it, which was called MindMeld and was shown at TechCrunch Disrupt in 2012. The app went away and the name stayed, and the company eventually renamed itself after it. Cisco announced its intent to buy the company on 11 May 2017 and closed the deal on 26 May for $125 million in cash.

Cisco then did the generous thing with it. The platform went open source, with the public repository created on 24 March 2019, and it ran for three more years as a real project with real users. By the end it had 694 stars and 193 forks, and the last release was 4.7.0rc2 on 1 November 2022. The homepage listed in the repository metadata is mindmeld.com, which today answers nothing on either scheme or either host.

~ the case file ~

~ what you actually had to build ~

The architecture diagram in that README explains its era better than any retrospective could. To make a machine understand "book me a table for four at eight", you first trained a domain classifier to work out which part of your product was being addressed. Then an intent classifier to decide what was being asked. Then entity recognition to find the four and the eight, with entity role labeling behind it to work out which number meant people and which meant time. Then entity resolution, to turn a restaurant name into a row in your database, and a parser to tie the pieces together. Every one of those was a separate model, and every one of them needed labeled training data you had to invent yourself, because nobody had ever said that sentence to your product before.

I build computer vision for edge hardware, so this pipeline is the shape of my working life in a neighboring domain. The pleasure of it was real. You could look at a misclassification and know precisely which of the six stages had failed and what to feed it. The cost was that a competent conversational product took a team a year, and the thing you got at the end only understood the sentences you had anticipated.

~ 28 july 2026 ~

What replaced all six stages was one call to a model that had never been shown your domain and did the job anyway. Cisco's README does not sulk about this, which is why it reads well. It grants that the platform was once state of the art, then tells the people still using it to go and find something better.

The useful date in that notice is not the archive date, it is November 2022. That is when the releases stopped, and it is also the month ChatGPT shipped. Cisco left the repository up for most of four years past the point where the work had effectively stopped, and the notice only got written when somebody finally did the housekeeping.

~ the forks ~

An archived GitHub repository is readable forever and changeable by nobody, which makes it a better gravestone than a dead domain. The 193 forks are still out there, most of them abandoned years before the parent was, and any one of them can still be cloned and installed by anybody who wants to see how this was done before the models got big. The documentation it shipped was unusually good, and it now describes a method rather than a product.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »