Two processes. No chat history. One verified resume point.
Session A performs work, verifies the result and records it with a signed work receipt. Session B starts as a separate process and resumes from that evidence-backed state.
What the proof demonstrates
Not semantic preference memory. MAGLA does not claim that a retrieved sentence became true because a model said it.
Verified work state. The application records an observed result after verification, with a bounded condition, evidence refs and a receipt tied to the task session.
Bounded failure memory. Failed approaches can be recorded without turning them into universal rules.
Independent resume. A new process recovers the last verified outcome and next action.
Reproduce it
Run the public 5-minute quickstart. The regression test also runs Session A and Session B as separate Python processes against a fresh isolated MAGLA gateway.
SESSION_A_SAVED=1 READBACK_VERIFIED=true SESSION_B_RECOVERED=1 REMEMBERED=Acme prefers annual billing and requires SSO. NEXT=Draft a proposal with annual billing and SSO
Category boundary
User/profile memory answers “what do we know about this user or world?” MAGLA continuity answers “what work was actually done, what failed under which conditions, and where should the next agent resume?” You can use both in the same agent stack.