A music startup has launched a Kickstarter campaign for a hardware instrument that does something unusual with AI audio models. Rather than generating polished, radio-ready tracks, Engram deliberately mangles input audio and produces glitchy, unpredictable artifacts. The company calls it a "field recorder for latent space" and frames it as an exercise in circuit-bending tiny AI models.
What Engram Actually Does
Engram is a sampler and groovebox built by Thoughtful Things, a music startup founded by Evan King. The device takes incoming audio, processes it through a small neural network running entirely on the hardware, and outputs something that retains fragments of the original input while warping it into unfamiliar territory. In a campaign demonstration, King asked the device for "piano" using his voice and received back a glitchy, vaguely piano-like sound that was recognizably derived from the request but far from a realistic piano sample.
This is not a Suno-style text-to-music pipeline where a user types a prompt and receives a finished song. Engram is designed for experimentation. The company's explicit positioning is that it pushes AI audio models beyond their trained boundaries, treating hallucinations and artifacts as creative features rather than failures to be eliminated.
On-Device AI With Training Transparency
Engram operates without an internet connection. The AI model runs locally on the hardware, which is significant for both privacy and creative intent. A cloud-connected sampler would process audio externally and return results, but an offline device means the latency stays in the musician's hands and the audio never leaves the box.
The model was designed in-house and custom-trained by Thoughtful Things. The company has committed publicly to training only on open datasets containing audio licensed for commercial use, such as CC-BY material. It has also stated it will never train its models on non-commercial, pirated, or otherwise stolen data. In an era where the provenance of training data is increasingly scrutinized, that transparency is a deliberate differentiator.
Open Firmware and Circuit-Bending Philosophy
Thoughtful Things plans to open-source Engram's firmware, allowing third parties to modify it or load their own custom-trained models onto the device. This extends the instrument's lifespan beyond a single company's creative vision and turns it into a platform for audio experimentation.
The company draws explicit inspiration from circuit bending, the practice of modifying electronic devices to produce unexpected and often chaotic sounds. Engram is built around the idea that the most interesting audio comes from pushing a model past its intended behavior, not from keeping it within well-behaved boundaries. The tiny AI models inside the device are designed to be tweaked, broken, and recombined.
Availability and Pricing
The Kickstarter campaign launched with a limited run of devices priced at $675, which represents a 30 percent discount from the anticipated retail price. Thoughtful Things has not finalized the retail price but has indicated it will likely land between $850 and $900. The company has not disclosed shipping timelines or total production quantities beyond describing the run as limited.
The Broader Context
Engram arrives at a moment when the music technology market is saturated with AI-powered tools that promise to simplify creation. Most of those tools optimize for accessibility and polish, guiding users toward familiar outcomes. Engram occupies the opposite end of that spectrum. It assumes the user already understands audio synthesis and is looking for a device that produces surprising, unconventional results rather than reliable ones.
The decision to keep the model small and local rather than cloud-dependent reflects a broader tension in AI hardware: whether the most creative outcomes come from the largest models with the most data, or from small models pushed to their limits through deliberate misuse and creative exploration. Engram is betting on the latter.