KAI MEMORY SWEEPTASK: Sweep today's inputs into Kai's memory — promote what's verified, reject what's unsafe, reflect, then re-assemble context
◆ DRIVE MEMORY BOOT (read-only · cached by file ID + modified time) ⠋ BOOTING
· BOOT_SEQUENCE.md · MEMORY_INDEX.md · semantic/context_canon.md · procedural/STANDING_RULES.md · KAI_RUNTIME.md · RAG_TRIGGER_INDEX.md
auto boot budget 0 / 12,000 chars · explicit "boot memory" gets 90,000
◆ MEMORY GATE only user-stated facts, decisions, preferences and tool-backed outcomes are promoted
INPUT (untrusted)VERDICTOUTPUT (where it lands)
◆ REFLECTION · wisdom.py new episodes since cursor 16/20 waiting for 20
episodes decay beyond 4,000 · cursor #18,431 · failed or unverified work never reaches reflection
◆ CONTEXT ASSEMBLY next turn: "build a lander for acme roofing"
prompt budget 0 / 12,000 chars · client memory never leaks into another client's turn