SCENARIO TIMELINE
Press ▶ JUDGE DEMO to run the Track-1 proof
scenario. Each step turns green only when the engine state verifies it.
01
TEACHdurable memories extracted from chat
02
SESSION BOUNDARYno raw history crosses over
03
CROSS-SESSION RECALLconstraints honored, never re-told
04
BELIEF REVISIONstale fact superseded, not duplicated
05
SLEEP CYCLEfragments consolidated, weak traces forgotten
● PERSISTENT MEMORY · CROSS-SESSION · SELF-ORGANIZING
An agent that never
forgets who you are.
ENGRAM distills every conversation into durable memories — deduplicated, reinforced, superseded when you change, consolidated while it sleeps. Watch the constellation grow.
~2 min, in an isolated fresh universe: teach → cross-session recall → belief revision → sleep consolidation. Every check is verified against the live store.
01 · TEACHIntroduce yourself — diet, work, style. Watch memories crystallize.
02 · RECALLAsk for a recipe — it respects your allergies without being told.
03 · EVOLVEChange jobs. The old belief is superseded, not duplicated.
NOT JUST CHAT / RAG
✓ Qwen distills chat into durable memories
✓ Hybrid recall — semantic + recency + importance + usage
✓ Arbiter supersedes stale facts, no contradictions
✓ Sleep cycle consolidates & forgets
✓ Every recall is scored — hover / click a recall chip
preference
semantic
episodic
procedural
MEMORY DECISION INSPECTOR
// send a message (or run the judge demo) — every retrieval
decision is dissected here: candidates, scores, rejections, the exact
context handed to Qwen, and the resulting memory operation.
0
Active memories
0
Reinforcements
0 tk
Memory context
0 tk
Full history
Memory trace
// no activity yet
Engram store
seeded demo loading — diet · allergies · deploy rules · cluster facts…