Engram Puts AI Sampling in an Offline Groovebox

Thoughtful Things has launched a Kickstarter campaign for Engram, its first instrument. The device is a sampler and groovebox that uses AI to mangle incoming audio and hallucinate new sounds. The Verge

Founder Evan King describes Engram as a "field recorder for latent space." Latent space is the internal map an AI model uses to organize what it has learned. Engram has no internet connection. It runs a tiny AI model locally, so all inference, the model's moment-to-moment processing, stays on the device itself.

Thoughtful Things says that model was designed in-house and custom-trained. The company states its audio models were trained on open datasets containing only audio licensed for commercial use such as CC-BY. It also states it has not trained and will never train its models on non-commercial, pirated, or stolen data.

For working musicians, that training history has practical weight. Sample clearance is still a constraint in commercial releases, and studios and labels now include training provenance in procurement checks. Engram sidesteps cloud dependency entirely, which also removes questions about latency, availability, and data retention that come with hosted generative audio tools.

Thoughtful Things plans to open Engram's firmware so others can tweak it or load custom models. The intent is to make the embedded engine itself open to modification, not just the sounds stored in it. In practice, that could mean patch behavior, model weights, and control mapping become editable layers rather than fixed firmware.

On function, Engram is built around capture and transformation. It can warp, blend, and extend user recordings in unpredictable ways. Thoughtful Things The company frames the instrument around a tweakable tiny AI engine for happy accidents of sound, using stochastic output, or random variation, as input for writing rather than as a mastering step.

The Kickstarter launch is structured as a limited run with pricing starting at $675. That $675 pledge is listed as a 30 percent discount, with a limited number of discounted units offered to early backers. Kickstarter lists the project as "Engram: Generative audio sampler & groovebox" by creator Evan King, categorized under Sound with a location of Los Angeles, CA. Its listing copy describes Engram as "Generate and mangle uncanny audio with a sampler powered by embedded audio models." Kickstarter A discovery listing showed 30 days left and 59% funded, within a search result of 12 projects for the term.

The broader context here is the shift of generative audio from workstation plugins and cloud APIs into dedicated, offline hardware. Grooveboxes already combine sampling, sequencing, and resampling in a loop built for performance. Adding local latent-space traversal to that loop changes the interaction. The musician no longer selects from a static library or waits on a render, but steers a small model in real time and commits the results to the sequencer.

In my view, the two decisions worth watching are the offline constraint and the open firmware plan. Keeping a custom-trained model fully embedded forces tradeoffs around model size, memory bandwidth, and control resolution, but it buys reliable behavior on stage and in the studio. Opening the firmware then invites the kind of third-party model porting and parameter hacking that sustained earlier sampler and modular ecosystems. If that combination holds, Engram could find durable use less as a novelty generator and more as a capture device for material that would be difficult to synthesize by other means.