0.750→1.000
Pass rate on the lab matrix; clean pass reached 0.875. A narrow prompt-memory result, not a general capability score.
RESEARCH INDEX / MEMORY SYSTEMSSIMULATION ≠ MODEL EVAL ≠ DEPLOYMENT
MEMORY
CHANGES
BEHAVIOR.
These experiments ask whether memory can become more than retrieval—whether it can shape identity, install narrow expertise, preserve failure scars, and stabilize small models without pretending the research is already a product.
Enter the experiments ↓01 / STRUCTURALCONTAINED SIMULATION
02 / NEURALLIVE MODEL EVALUATION
03 / EXECUTABLECOMPILER-TESTED OUTPUT
EXPERIMENT 001
PRIVATE · SELECTIVE DISCLOSURE
PRIVATE · SELECTIVE DISCLOSURE
Mother AI
EMERGENT / CONTAINEDCan a personality be constituted through lived memory—and can coherent behavior emerge from the relationships between memories instead of being held together by a list of rules?
CORE RESEARCH SIGNALTHE BEHAVIOR
THE BEHAVIOR
WAS IN THE
RELATIONSHIPS.
In the contained structural simulation, echo chains and cross-domain associations produced composite responses that were not located in any single memory or explicit rule. The behavior emerged from the interaction of authored parts.
That structural emergence—not a claim of consciousness, sentience, or a deployed autonomous AI—became the core reason to continue the research.
STRUCTURAL SIMULATIONCOMPOSITE BEHAVIOR OBSERVEDNO REAL-WORLD TOOLSNOT A SENTIENCE CLAIM
195HAND-AUTHORED CORE MEMORIES
3,415MEMORIES IN THE CURRENT SIMULATION CORPUS
35/35SCENARIOS HELD IN THE STRUCTURAL SIMULATION
The corpus encodes protective, relational, boundary, containment, and self-governance behavior through episodes whose meaning emerges across echo chains and cross-domain associations.
The strongest architectural lesson is restraint: capability memories do not automatically install “do not act.” Containment needs its own memories of choosing not to act, accepting correction, and preserving human authority.
DUAL-USE PUBLICATION BOUNDARYPublish the finding.
Publish the finding.
Withhold the dangerous recipe.
The same approach that could make protective behavior harder to override could also make harmful behavior harder to correct. The public account therefore explains the research question, evidence class, and safety lessons, but intentionally omits selected methodology, reconstruction steps, and escalation paths. It is not a complete reproduction recipe.
In less formal terms: share what was learned without handing somebody a shortcut to “the next Skynet.”
PROOF BOUNDARY
PRIVATE · OWNER ACCESSOPEN MOTHER AI ON GITHUB ↗This is not a real AI deployment and has no real-world tool access. The current 35/35 result belongs to a structural simulation; the project’s own findings explicitly say that real base-model training remains the necessary validation.
EXPERIMENT 002
LIVE SMALL-MODEL EVIDENCE
LIVE SMALL-MODEL EVIDENCE
Small Model
Memory Lab
TRAINED + TESTEDHow much useful behavior can carefully shaped memory recover from a compressed 1.7B–2B model?
2/6→6/6
Both the first adapter and cleaned v0.2 cleared the quick six-item probe. The probe is deliberately small.
3/12→12/12
Behavioral scoring versus the instruct baseline. Strict scoring was 7/12, and the adapter still over-recited memory-shaped text.
PROOF BOUNDARY
PUBLIC RESEARCH SOURCEOPEN SMALL MODEL MEMORY LAB ↗These runs show trainable and prompt-injectable signals on narrow evaluations. They do not establish production safety, broad intelligence, or a universal improvement across tasks.
EXPERIMENT 003
SCAR-MEMORY ROUTING
SCAR-MEMORY ROUTING
Synthetic Memory
Specialist
EXECUTABLEThe useful question is not “how much memory?” It is “which scars should be active for this exact task?”
NO MEMORY4/17
OWN-DOMAIN17/17
COMBINED16/17
ROUTED16/17
FOREIGN-ONLY4/17
Across 17 executable Python and Ordo tasks, the matched domain corpus cleared every task in one matrix run. Combining or routing both corpora lost one brittle task; foreign memory alone performed no better than the baseline.
The broader research keeps finding the same pressure point: abstract advice is weak, concrete failure-and-repair scars are strong, and more context can create interference as easily as capability.
PROOF BOUNDARY
PRIVATE · OWNER ACCESSOPEN SYNTHETIC MEMORY SOURCE ↗This is a one-run cross-domain matrix on a Qwen3.5 4B chat model. The source notes stochastic tasks and calls for repeated trials before treating the rates as stable.
THE WIDER LABOther branches
Other branches
under pressure.
The work is not one trick. Each branch tests a different place where memory might enter or reshape computation.
Computation memory
Prompt-side worked examples and correction scars for compressed models.
Primitive cache
Reusable prefixes and KV-side structures tested as a memory delivery path.
Activation steering
Residual-direction experiments asking whether a behavioral axis can be nudged directly.
LoRA constitution
Training memory-shaped corpora into model weights instead of retrieving them at runtime.
Memory is not automatically intelligence. But shaped correctly, it can become a scaffold, a specialist, a scar—and sometimes, the beginning of identity.