Persistent Memory

What it is

Persistent Memory is a local-first point-and-click adventure about generated dialogue governed by deterministic, inspectable, provenance-aware memory.

The Build Week story, The Tower's Terms, begins when Bramble accuses Daria of stealing the village's banked hearth ember. The player follows that claim through rumor, firsthand testimony, physical observation, and inherited continuity before deciding what Mallow's communal record should retain. Only then can the player act on the corrected record and wake the village aqueduct in amber light.

Why it is different

Most AI games let generated dialogue quietly become game truth. Persistent Memory draws a harder—and more interesting—boundary. A model may recognize intent, voice a character, synthesize admitted evidence, or propose an attributed memory. It cannot invent physical evidence, press the player's record seal, or mutate the world.

Every durable claim carries a stable identity, source, stance, confidence, and provenance chain. The authoritative state is an append-only semantic event log. Reloading reconstructs the same record and world from those events alone, regardless of which inference endpoint supplied the prose.

Bring Your Own Inference

Local inference is the intended experience, not a fallback. Players can use Ollama on the same machine, privately controlled hardware, any OpenAI-compatible Chat Completions endpoint, OpenAI Responses, or the deterministic offline floor. An optional loopback-only bridge also lets a developer test Luna, Terra, or Sol through an existing ChatGPT/Codex subscription without exposing credentials to the game.

Endpoint settings, credentials, prompts, and generated responses never enter the save. Switching the voice changes expression—not truth, legal progression, or save compatibility. The submitted one-take demo runs live on Qwen 3.6 35B A3B using hardware I control.

How Codex and GPT-5.6 were used

Codex was my primary Build Week development environment. Across one continuous core session, it accelerated the architecture, TypeScript/Tauri implementation, provenance and replay tests, lore integration, endpoint hardening, bespoke visual asset workflow, playtest fixes, and final submission preparation. Codex also built a deterministic judge replay that proves endpoint changes cannot alter the semantic event path or final snapshot.

GPT-5.6 Sol was invoked through the optional local subscription bridge to test the same provenance-aware dialogue path against a frontier model. GPT-5.6 Luna validated the strict structured-intent path. Those comparisons refined prompts and adapter boundaries while preserving the project's central promise: Persistent Memory is local-first and endpoint-agnostic.

Technical evidence

• Deterministic truth and generated language are separate layers.
• Immutable memories preserve claim identity and full provenance.
• Only authored game actions create semantic events.
• The Far Tower Advisor can synthesize admitted evidence but never writes memory.
• Ollama, Chat Completions, Responses, Codex subscription, and offline modes share the same save and legal route.
• The automated replay restores the final state solely from the append-only log.
• The complete route remains playable when a live endpoint fails.

The thesis is simple: The Tower may answer for us. It must not remember for us.

Built With

• codex
• css
• gpt-5.6
• html
• ollama
• openai
• qwen
• rust
• tauri
• typescript
• vite
• vitest