I run a text-based d20 RPG where an AI Game Master narrates the adventure, one chapter at a time. The classic wall: ~6,000 tokens of context, and the GM physically could not remember what happened at turn 20 when it hit turn 21.

I fought it with summaries — which then got summarized until the oldest ones were a summary of a summary of a summary, and anything specific from a dozen turns back was simply gone. Not fuzzy. Gone.

So I rebuilt the memory system on one decision: fold, never delete. When turns leave the visible context, their original text moves verbatim into a transcript instead of being destroyed. Then four layers on top:

Lorebook — a structured canon (NPCs, items, places, factions) plus a list of open threads, refreshed after every turn and re-read every turn. Crucially it’s owned by the player, not the model: it’s an editable panel, so the canon stays honest.

Archive summaries — folded chunks become structured EVENTS / FACTS / THREADS summaries with tags, injected into every prompt. The past gets compressed, never evicted.

Scene anchor — who’s present, where, when, what’s at stake. One fixed point to stop the “wait, where are we?” drift.

Tagged retrieval — a cheap matcher pulls verbatim chunks of the original text back into the prompt when the current scene matches. Actual scenes, not retellings.

Honest about the edges: retrieval is keyword matching, not understanding (turn 90 says “the key from the tomb,” turn 45 called it “the glimmering key” — recall can miss, though the lorebook usually catches it), recall is capped in characters, and the oldest canon still gets evicted past the hard caps. Net result: roughly 10× the memory horizon on the same 6k context, ~70–80% of what a perfect system could do.

The practical test: turn 60 referencing turn 3 now works. That used to be impossible, not merely unlikely.

Full writeup: Perchance for Humans

(Built entirely by prompting — I’m not a coder; the AI assistant wrote all of this for me.)