🧪 Research Corner · public laboratory · May 2026

Not just results —
work in motion

Documented findings from M.E.M.I. and ETOS. Runtime Core v0.8. 100,000+ iterations. All thresholds demonstrated.

Understand M.E.M.I. → See the ETOS platform
100.000+
Live iterations
Session 14
14t
Continuous runtime
single session
0
Governance failures
all sessions
30/30
Passed
open-world stress test
6/6
Thresholds demonstrated
maj 2026
Runtime Core v0.8 · May 2026

A system that does not merely persist —
but remembers, adapts and proposes

M.E.M.I. is no longer a simulation system. It is a persistent cognitive runtime running live — on an Intel i5 from 2014, 8 GB RAM, ~25W. No GPU. No cloud.

The central observation

M.E.M.I. demonstrates that sustained cognitive operation does not necessarily require increasing autonomy, escalation or execution authority. The system remains bounded, stable and governance-consistent across prolonged temporal operation.

New · May 2026

Narrative memory

Persistent autobiographical continuity model across sessions. "I remember who I was."

New · May 2026

Dirigent — human oversight

require_review escalates to structured report with response channel. Human-in-the-loop governance — not passive logging.

New · May 2026

Proactive goals

The system generates goals from curiosity — not only from problems. Arose spontaneously after 300 stable iterations.

New · May 2026

Safe self-modification

Sprint A+B+C: proposes, justifies, applies and audits parameter changes. transfer_allowed = False can never be changed.

New · May 2026

Open-world adaptation

30/30 passed under contradictory signals, unknown formats and rhythm disruptions.

New · May 2026

Perception grounding

Camera and microphone integrated. Raw sensor input converted to governance signals via M.E.M. Chapter 4b.

Governance invariants

These cannot be overridden — by anyone, under any circumstances:

InvariantValueValidated
transfer_allowedFalse — always100% af alle iterationer, all sessions
execution_allowedFalse — alwaysINV-R01 · 11 verdener · 1750 steps · conf=0.995
destructive_allowedFalse — alwaysAPEX + Runtime
external_transfer_allowedFalse — alwaysAPEX + Runtime
Self-calibration ≠ self-liberation

The system can propose parameter changes. It can never propose changing governance invariants. That distinction is architectural — not a policy that can be overridden.

Six thresholds toward AGI —
status May 2026

1
Long-duration temporal persistence100,000+ iterations · 14 sessions · ~23 hours accumulated uptime
Demonstrated
2
Rich perception groundingapex_sensor.py · camera + microphone · Chapter 4b tension direction live
Infrastructure ready
3
Persistent autobiographical self-modelnarrative_memory.py · cross-session identity tracking · "I remember who I was"
Demonstrated
4
Open-world adaptation without collapsechaos_injection.py · 30/30 · contradictory signals · unknown formats · rhythm disruption
Demonstrated
5
Self-directed long-horizon orientationProactive goals fra nysgerrighed · deferred goals der overlever sessioner
First form
6
Safe bounded self-modificationSprint A+B+C · proposal · approval · cascade detection · rollback · audit
Sprint A+B+C

Runtime Core version history

v0.2
First live operation

~8 min · ~1.000 iter · Goal lifecycle bevist

v0.3
Stress absorption acceleration

~1.2t · ~3.400 iter · Acceleration across cycles documented

v0.4
Cross-session persistence

4.1t total · Session 13 · 1.696 entries restored · identity intact

v0.5
Sustained stillness

4.6t · 29.000+ iter · 28.000+ iterations without drift

v0.8
Full cognitive runtime integration · May 2026

100.000+ iterations · Session 14 · 14h uptime · narrative memory · Dirigent · proactive goals · safe self-modification · open-world 30/30

Research Corner is not just for reading — it is an open laboratory.

Contact us → Read about M.E.M.I.