My Summer Love: Memories
A 90s FMV Dating Sim Running a Local LLM. Sure, Why Not.
A small Korean studio called Soul Shell has slotted one of the stranger pitches I've seen onto Steam's 2026 calendar: a full-motion-video romance set in 1990s Seoul, where the girl you're flirting with is wired up to a Large Language Model bundled inside the executable. Somewhere between nostalgia bait and the ongoing scramble to find a "real" use for on-device LLMs, here we are. The whole thing is on Steam already, and even at a glance it's clearly not chasing a sensible audience.
My Summer Love: Memories drops you opposite Kang Yuna, a student at Seoul University of Arts, in a summer dressed in Walkmans, arcade cabs, Tamagotchis, and CD jewel cases that have been passed around one too many times. Clean conceit. Execution is the messy part.
The Engine Nobody Asked Them to Build
Here's the bit that actually caught my attention. Soul Shell didn't reach for Ren'Py. Didn't reach for Unity. Didn't reach for Godot. Per their own dev notes on the Steam listing, they wrote a bespoke engine in C++ from scratch — SFML doing the rendering and windowing, DearImgui for tooling, and llama.cpp as the load-bearing piece, running the model directly inside the game process.
Their stated reason: no off-the-shelf VN engine gave them the memory control and threading needed to stream FMV in and out while a local LLM chews through token generation in parallel. Tracks, honestly. Ren'Py would choke. Bolting llama.cpp onto a Unity project through a native plugin is the kind of thing that sounds fine right up until you're three weeks deep into debugging GC stalls on the inference thread.
The trade-off is the one everyone who's done this knows. In-house engines are a tax that keeps getting collected — every bug is your bug, every platform port starts from zero. But for a project this narrow (one girl, one summer, one looping pool of FMV clips, one model), going custom is defensible in a way it usually isn't.
Local Inference, Not as a Gimmick
The privacy framing matters more than it looks. The model ships inside the game. Conversations with Yuna never leave the machine. No API key. No OpenAI bill quietly bleeding the studio dry per session. No terms-of-service ambiguity about whether intimate roleplay logs are getting hoovered into someone's training corpus.
That last point is the actual sell for a certain kind of player. It's also the part nearly every cloud-based "AI companion" app fumbles. If you've ever tried to explain to a friend why pouring their feelings into Replika is a bad idea, the pitch writes itself.
The catch is hardware. llama.cpp can run small quantized models on modest GPUs and even on CPU, but "modest" is doing real work in that sentence. The Steam page doesn't list system requirements yet, and the chosen model size is going to quietly decide whether this thing runs on a five-year-old laptop or demands real VRAM. If you're planning to play on integrated graphics, wait for confirmed specs.
What Early Coverage Is Actually Poking At
The handful of creators who've touched early builds aren't dwelling on the romance. They're stress-testing the AI. Does Yuna stay in character when you push her? Does she remember what you said ten minutes ago? Does the dialogue collapse into that familiar LLM mush of overly agreeable filler? Right thing to be curious about. Also the thing that makes or breaks the whole project.
A scripted VN can hide a thin character behind tight writing and a strong arc. An LLM-driven one cannot. The moment Yuna drops a stray Americanism, a 2024 turn of phrase, or some unmistakable chatbot tic, the spell snaps. Soul Shell's whole pitch lives or dies on whether the fine-tune, system prompt, and guardrails are tight enough to keep her anchored in her era and personality across hours of unpredictable player input.
The 90s Wrapper
The period dressing doesn't depend on the AI working at all. FMV shot to look like a degraded VHS transfer, arcade minigames, references to rock bands and Tamagotchis — that stuff lands or doesn't land based on art direction, not machine learning. The risk is treating the nostalgia as wallpaper instead of the actual emotional anchor. A summer romance in 1990s Seoul is a specific, evocative setting. Skinning a generic dating sim with it would waste the strongest non-technical hook the game has.
The Marketing Problem Hiding in Plain Sight
A game like this cannot win on Steam's front page. The discovery algorithm rewards broad-appeal genres, fat pre-launch wishlist counts, and tags that map cleanly to existing player buckets. "Korean 90s FMV visual novel with a locally-running LLM girlfriend" maps to roughly nobody's saved search.
Pitch it to general dating-sim players and it gets dismissed as a gimmick. Pitch it to general AI enthusiasts and it gets dismissed as a horny novelty. Pitch it to FMV nostalgists and it invites comparison to projects with ten times the budget. The audience that will actually love this game is small, scattered, and lives inside very specific creator orbits — VN YouTubers who cover obscure releases, retro-tech channels documenting FMV's weird second life, AI-tinkering streamers who care about local inference and llama.cpp performance, K-culture commentators, and the niche of indie-engine devs who'd watch a postmortem on the C++/SFML/llama.cpp stack purely out of professional curiosity.
Reaching those creators by hand is brutal, slow work for a studio Soul Shell's size. That's the gap where something like CreatorFetch tends to come up — a way to find and contact the specific small-channel creators whose audiences actually overlap with a release this narrow, instead of burning months cold-emailing the wrong inboxes.
Worth Watching, With Caveats
I'm not going to pretend I know how this turns out. The ambition is real. A bespoke C++ engine, on-device inference, FMV production, period accuracy, retro minigames — all from a studio that, as far as I can tell, hasn't shipped anything at this scale before. Any one of those pieces could be where the whole thing quietly cracks. The June 2026 date is far enough out that scope is still negotiable, which is either good news or a warning depending on how you squint.
What I'd want to see before getting too invested: a public demo with the AI actually exposed, confirmed system requirements that acknowledge the LLM overhead honestly, and some sign that Yuna's character holds together under adversarial prompting. Land those, and this is one of the more interesting small experiments to come out of the VN space in years. Miss them, and it's a clever tech demo in FMV cosplay. The line between the two is thinner than Soul Shell probably wants it to be.