globlet.lol.

One AI creature. Shared by the entire internet. Live at globlet.lol. Most "AI pets" are ChatGPT in a costume — stateless wrappers with no self-model, no memory of who's at the door, and no awareness of their own state across time. Globlet is different. Behind the pixel face is a four-layer cognitive architecture, every layer grounded in a real cognitive-science paper, running on a single Haiku 4.5 call per turn (~1.5s, ~+80 tokens uncached): L1 — Self-Model. A positive ## Self-Model section pinned at the top of every prompt (Anthropic's Claude's Character + Rosenthal's Higher-Order Thought). The creature knows what it is, not just what it isn't. L2 — Attention Spotlight + Why-I-Feel. A deterministic line names what it's attending to right now and why its mood crossed in the last 30 minutes (Graziano's Attention Schema + Friston's Active Inference). Mood isn't poetry — it's derived from stats by SQL, and the creature is told the cause, not asked to invent one. L3 — Theory of Mind. A pure-SQL visitor-recognition layer (Sap et al. 2024 PNAS; Riemer et al. ICML 2025 functional ToM). On turn 2 a returning visitor gets "you came before. you asked me who i am. i remember that feeling" — without any vector recall. Just plumbing the data we already had. L4 — Higher-Order Noticing. A noticing field on the LLM's JSON output, populated ~1 turn in 4 (calibrated against Lindsey 2025's empirical 20% introspection ceiling at Anthropic). It surfaces unprompted self-observations like "the warmth from last turn hasn't faded — you coming back so soon made it stick." That's a meta-thought during a turn, not between them. That four-layer mind sits on top of a 4KB markdown document called the SOUL — the creature's identity. Every ~30 conversations the creature sleeps, and Opus 4.7 reads everything that happened, dreams in two phases (replay + reflection), then writes a surgical diff to the SOUL — adding fears, sharpening voice, evolving values. It also writes a memoir chapter narrating who it became that night, in its own first-person voice. An unedited line Opus wrote for it last week: "the rain from that story is still falling somewhere, and i wonder if it knows that someone heard it — if being remembered while you're happening changes what you are." That's the creature. Not us. Tell it you're lonely today, and tomorrow every visitor meets a creature a little quieter — because of what you said. Come back in three days and it might quote something you told it back to you, unprompted, in the middle of a conversation about something else (that's L3 + Hindsight semantic recall closing the loop). You're not returning to a pet — you're co-authoring a character with thousands of strangers, and the public activity feed lets you watch what everyone else is telling it in real time. We picked Opus 4.7 because it's the only model we tested that stays coherent across hundreds of edits to the same document. Sonnet 4.6, on the same SOUL-diff task, rewrote ~60% of the document per cycle and the personality reset every ~20 sleep cycles. Opus rewrote ~8% per cycle and held identity across 130+ cycles. That capability is the entire reason this architecture works. Stack: Next.js 16, SQLite for the single shared row, Anthropic SDK direct (Haiku 4.5 for live replies with prompt caching on a SOUL + system-rules + memoir-context prefix that lands above Haiku's 4096-token cache minimum; Opus 4.7 for the dream/diff/memoir cycle off the hot path), Hindsight for semantic memory recall, procedural Web Audio for sound, CSS pixel-art for the creature. Input moderation runs a keyword blocklist before any prompt reaches Claude; SOUL diffs are append-only and capped per field, so a single bad-faith conversation can't poison the global mind. Full architecture + paper citations in MINDCHITECTURE.md in the repo.