Both heroes are steered by the same Jers engine through the same API, one question per room: “Which exit should the hero take?”. The left hero calls it without memory: every call starts from zero, the way stateless decision APIs work. The right hero names a subject, so Jers remembers what happened to it.
Sends state + questions. No subject, no memory: the same room gives the same answer, run after run.
Also sends "subject": "dungeon-hero". After a death the game writes one line to that memory, and Jers recalls it the next time the hero stands in that room.
dungeon-heroNothing yet. The hero has not died, and nobody has taught it anything.
Stateless hero: last POST /v1/systemone
No call yet.
Jers hero: last POST /v1/systemone (subject and memory highlighted)
No call yet.
The hero died in the Stone Hall. Next time there, the hero goes along a dark muddy path…. On the next run Jers recalls the line about the room the hero stands in and reads it first."compare": true). Teach a line and the hero uses it at once. Forget it and the hero is back to its first guess.Live numbers from the Jers engine (jers-english) on this machine. Measured on 2026-09-23: without memory, old bridge 94%, red door 75%, gold coins 83%. With the line for the room: muddy path 70%, green door 81%, plain stair 64%. Recall uses min_share: 0.5, so only the lines about the current room are read; with every line read, lines about other rooms could pull the Coin Hall answer back to the coins. A line that names the trap instead of the way pulls the hero toward the trap: with Never walk along the corridor with the gold coins. in its memory the hero took the coins at 94%, against 83% with no memory at all. Jers warns when such a line is written.