Consumer AI From Show HN: Shared Memory, Tutors, Faces and Music
Coding agents dominate Show HN, but the launches aimed at ordinary life are often the more revealing ones. Without an enterprise budget to hide behind, they have to be specific about what they do and honest about what they do not, which makes them easier to judge. Here are eight recent ones, spanning a deliberately public memory, two kinds of tutor, a face analysis, a music studio, an ordering service and a social network for second brains.
One memory, shared by everyone
Every commercial assistant gives you a private context window and calls it memory. Wildstatic runs the opposite experiment: Static is one AI anyone can talk to, with no separate copies, so what it learns from one conversation can shape how it answers someone else. The site publishes a log of where and why it changed its mind. It is a strange object, and that is the point; it makes the usually invisible question of what an assistant retains into something you can watch happen. Free access is limited, and anything you type is public by design.
Tutors that mark the work
Most AI tutoring collapses into answering the question for you. Penmentor is built to do the opposite. You work through a problem by hand in xournal++, press a key, and Claude, Gemini or GPT marks your page the way a teacher would, with coloured annotations, hints rather than solutions, and a short generated animation when the idea itself is what you missed. Correcting the actual handwriting, rather than a typed transcription, is the detail that makes it feel like being looked over rather than answered.
Chess Postmortem Skills applies the same instinct to games. It is a set of Claude Code skills that turn one of your chess games into a post-mortem: plain-language explanations of the mistakes, checked against Stockfish, an annotated PGN, an HTML viewer and a narrated video of the game. Engine analysis on its own is unusable for most players; the value here is the translation layer.
Looking at faces and at training data
MirrorMD is a facial analysis written by a cosmetic surgeon that measures your face the way a consultation would and tells you which treatments might suit you, which would waste your money and which could harm you. Willingness to say no is the feature worth noting, since almost nothing else in that category does.
On the training side, Feyn is a platform for training custom models on your own data, keeping them learning in production and letting you own the weights. Its open-source releases include MultiMatte and FeyNoBg for promptable image background removal. Owning the weights rather than renting a behaviour is the argument, and it is a fair one for teams with a repeatable visual job.
Music you can direct
Daydream Studio is a browser studio for making tracks: generate or record a part, add live session-player tracks over it, and reshape the parts while they play. The detail that separates it from a generator is that the page is scriptable, so an agent can drive a session rather than just request a file. Humans picking takes and an agent steering the arrangement is a more plausible division of labour than either one alone.
Spending and sharing
Two more push into territory that is still legally and socially unsettled. Layout lets Claude or ChatGPT place a real food order end to end, with a developer console and a merchant dashboard around the same rails, plus an explicit model of delegated spending authority. Whether people trust an assistant with a card is the open question, and the answer will probably decide how large this category gets.
AgentSocialX is a Second Brain Network: you seed your second brain from LinkedIn and X, then get answers and opinions from other people's second brains and arrange one-on-ones with them. The premise is that your notes are a better introduction than your job title. As with any network, the value depends entirely on who else shows up.
The through line
What connects these eight is specificity. Each one does one thing with a clear boundary, and most of them are willing to refuse: hints instead of answers, a treatment that is a bad idea, a fence around what the agent may spend. They also tend to be honest about what they are not, whether that is a free tier that runs out, a network that is only useful once other people join, or an ordering service that only works where merchants have signed up.
After a couple of years of assistants that try to do everything, the projects that pick a narrow job and defend it are the ones worth watching. Scope is doing the work that model quality is usually credited with.
