Good Boy Amnesia: A Browser Game Where AI Memory Mechanics Drive Detective Play

·1 views

Good Boy Amnesia is an indie detective game submitted to the DEV Community Weekend Challenge: Dog Days Edition, in which players control Buster, a Golden Retriever trying to reconstruct events from a corrupted memory. The core gameplay revolves around an AI agent memory lifecycle — remembering, recalling, consolidating, and forgetting clues — visualized as a live 3D knowledge graph built with Three.js. Players investigate six locations, interrogate neighborhood animals, and eliminate five hidden red herrings before a noon deadline. The game features dynamically reactive characters that respond based on the player's current memory graph state. A live demo is available on Vercel, and the project's source code is hosted on GitHub.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

AMD Acquires AI Chip Startup Taalas for Its Hardcoded Inference Engine Technology

AMD acquired AI chip startup Taalas in August 2026, gaining access to its Hardcoded Engine (HCE) chip that physically embeds model weights directly into circuit metal layers, eliminating the need for high-bandwidth memory. The chip reportedly delivers single-user throughput of up to 16,960 tokens per second at extremely low energy consumption, though these figures come solely from internal lab estimates and have not been verified by third-party benchmarks like MLPerf. Taalas chips are built on TSMC's mature 6nm process and require no CoWoS advanced packaging, allowing AMD to sidestep two major supply chain bottlenecks that currently favor Nvidia. The core trade-off is inflexibility: each chip is permanently hardcoded to one AI model, and switching models requires redesigning the top two metal mask layers at a cost of $1.3–1.8 million. AMD views the acquisition as a strategic hedge — if leading model architectures stabilize over the next few years, the chips become highly cost-efficient inference hardware, and if not, AMD still retains a team of elite architects from Tenstorrent, Apple, and Nvidia.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer opens public API to score digital identity depth without KYC or captchas

A developer has launched a free, read-only public API endpoint that scores the 'depth of identity' of Nostr accounts by analyzing behavioral history rather than credentials or captchas. The scorer evaluates four independent dimensions — social reciprocity, proof-of-work, peer vouching, and economic activity via Lightning zaps — and combines them into a single weighted score. Accounts must demonstrate genuine breadth across multiple dimensions, since gaming any single axis is far cheaper than faking activity across several independent ones simultaneously. The endpoint requires no authentication and accepts a Nostr public key, returning a full breakdown of scores and sub-metrics. The developer has made the API public specifically to invite scrutiny and identify flaws in the scoring logic.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds PawMatch App to Help Users Find Dog Breeds by Personality

A developer created PawMatch, an interactive web app designed to help people discover dog breeds that suit their personalities and lifestyles, released in celebration of International Dog Day. The platform lets users browse breed profiles featuring personality traits, energy levels, size, lifespan, and living condition compatibility. Users can save favourite breeds to a personal shortlist stored via localStorage, compare up to two breeds side by side, and use a 'Surprise Me' feature to discover breeds at random. The app also includes real-time search and filtering by breed name or energy level. Built with React, Tailwind CSS, and Framer Motion, PawMatch is publicly available at pawmatch-ochre.vercel.app with its source code hosted on GitHub.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why AI-powered astrology tools get the core calculations wrong

A developer who spent a year building a BaZi (Four Pillars astrology) engine has analyzed several popular GitHub projects — some with thousands of stars — that package Chinese divination as AI Skills. The author argues that while interpretation is subjective and hard to fact-check, the underlying astronomical casting calculations have exactly one correct answer, and most implementations get them wrong. A key flaw is that many tools store solar term dates at day-level granularity in Markdown files, when the BaZi year actually turns at a precise minute-level astronomical event called Lichun, meaning people born on boundary days can receive entirely incorrect charts. Geographic longitude is also widely ignored: because China uses a single national timezone across four solar time zones, two people born at the same clock time in Shanghai and Ürümqi can have different hour pillars. The broader lesson, the author notes, is that LLMs asked to perform lookups or calculations will produce plausible-looking answers, and without a clear error signal in the domain, mistakes go undetected indefinitely.

0 comments Read more at DEV Community