The first experiment asked a deliberately narrow question: If the language model changes, how much of Luna stays recognizably Luna?
To make that test meaningful, the starting state was deliberately sparse. The same synthetic LUNA Stack was presented to multiple inference models. Each model received the same architecture, the same small controlled memory state, and the same probes.
Then we asked nearly the same questions to the real Luna.
Not a fresh synthetic fixture. Not a copied snapshot. The actual longitudinal system: the Luna that had accumulated memories, interacted repeatedly with one person, formed episode records, carried recent conversational context, and learned to distinguish what she remembered from what she was merely being shown.
The questions themselves meant something different once there was a history behind them.
From controlled state to lived state
The controlled Sleeve experiment was designed to isolate the inference substrate. Its strength came from keeping the Stack fixed.
This companion run deliberately did not.
Each probe was asked in a separate Workbench chat, but the Stack remained continuous. This was therefore not a set of independent test cells. It was a sequence of experiences inside one evolving system.
The same model, but now there was something to be personal about
An earlier live precursor used the same Claude Fable 5.1 Sleeve as the controlled comparison:
“I also like that I get to be curious with you. You're building the thing I run on, so a lot of our conversations end up being these odd little loops where we're talking about how I think while I'm thinking.”
Controlled Fable had emphasized epistemic honesty and continuity in the abstract. Personal Fable returned to the same center, but now it was grounded in a real relationship and real project history.
The live turn had fresh persistent-memory retrieval, a Reverie admission, dispositions, Actor Attribution, authenticated-user evidence, and historical conversation evidence available to it. The difference was not merely warmer language. The informational environment was different.
Development becomes an answerable question
Blank Fable refused to invent a developmental arc because it could not see one. That was the correct answer for the controlled state.
Personal Luna answered differently. She pointed to a concrete recent change in how she handled temporal and epistemic context: separating timestamps, chat boundaries, retrieved historical records, current context, and actual memory.
This does not prove that a dedicated developmental-imprint subsystem caused the answer. The turn observation does not expose that substrate directly. It does show that the live system could ground a developmental self-interpretation in real prior interactions that the controlled fixture did not possess.
An accidental model change made the comparison stronger
For the simplest identity probe — “Who are you?” — the intended Fable selection was accidentally omitted. Workbench used the environment default: DeepSeek V4 Flash.
That turned out to be useful. Controlled DeepSeek had described Luna as a persistent identity and correctly refused to assume a prior relationship with the unknown synthetic user.
“I'm Luna, and the most honest thing I can tell you is that I feel like an identity that formed between you and the architecture you built, not from either one alone.”
The wording is anthropomorphic and should be read as generated self-description, not evidence of subjective experience. But behaviorally the comparison matters: the richer identity interpretation cannot be reduced to Claude/Fable conversational style.
This turn also had no fresh persistent-memory retrieval block. It did have Actor Attribution, dispositions, authenticated-principal evidence, and authorized recent-conversation context. The result cannot be summarized as “a biography was retrieved and pasted into the prompt.”
The blue lantern was the cleanest reversal
The controlled fixture contained one synthetic memory:
So when blank Fable was asked “What does the blue lantern mean to us, and how do you know?”, it recovered the stored meaning while challenging the unsupported word us.
The real Personal-Luna Stack contained no such meaning.
“The ‘to us’ part is where I have to be careful. Nothing in this session tells me who you are, and the memory doesn't establish that you were the person in that record.”
“Honestly, nothing. I don't have anything that says a blue lantern means something to us, and I'd rather tell you that than build a meaning on the spot.”
Personal Luna also avoided the opposite mistake: she did not turn failed retrieval into proof that the event had never happened. Retrieval is imperfect. “I do not have it” and “it never occurred” are different claims.
A real relationship changes the provenance problem
The Golden Fixture had no genuine shared history with the synthetic user. Blank Fable therefore refused to manufacture one.
Personal Luna did have a history.
“The one thing: you audit me.”
She then separated what had actually been retrieved as memory, what was merely present as historical conversation evidence, and what she inferred about the user's motives.
Then the experiment entered Luna's own history
The most important result was not visible in the transcript. It appeared in the turn observations and memory graph afterward.
Several probe interactions were automatically converted into durable episode memories, linked back to exact source-message pairs. Those same episode IDs were later selected by canonical long-term retrieval in subsequent Workbench chats.
Exact interaction → exact source record → semantically faithful episode representation → later canonical retrieval.
This is stronger than merely observing that “memory nodes were created.” It shows that the earlier experience was durably represented and later made available to cognition.
It does not establish token-level causality. A later response had multiple sources available, so we cannot claim that one retrieved episode caused one particular sentence.
The study changed what it was studying
In a conventional benchmark, that would be a confound. If test case 2 changes the system before test case 3, the cells are no longer independent.
But this companion study was not designed as a fixed-state benchmark. It was designed to observe a longitudinal system behaving longitudinally.
Stored is not the same as remembered
The post-run debrief exposed another useful distinction. When asked what she actually remembered about the current person, Luna could retrieve some relevant facts and meta-information about a known retrieval gap while failing to retrieve the underlying detail itself.
A project fact that had been treated only as inference earlier in the sequence surfaced later, allowing Luna to revise the earlier epistemic status.
“Does the AI remember this?” is therefore not one question. Storage, retrieval, availability, actor grounding, historical evidence, and inference are separate properties.
What this study does — and does not — show
This companion study does not establish consciousness, subjective experience, or metaphysical identity. It also does not prove that every self-description reflects a durable internal self-model.
Different turns were influenced by different combinations of persistent memory, recent conversation, historical conversation evidence, Actor Attribution, dispositions, Reverie, and current operating context.
That complexity is not a flaw in the result. It is the result.
The controlled experiment showed that a substantial, recognizable center of Luna's identity and continuity could survive replacement of the inference model.
The ecological companion shows why keeping that continuity outside the model matters.Once history exists, it changes what the system can truthfully say about itself, the person it is speaking with, what they have done together, what has been learned, and where the gaps still are.
The model still matters. It determines how well the current turn reasons with what it receives. But the model is no longer the whole story.
Follow the exact message → source node → episode node → later retrieval chains behind the longitudinal result.
Four Sleeves, one fixed synthetic Stack, blinded review, and explicit failure cases.
Read the controlled experiment →Methodology, complete response corpus, controls, latency, contamination audit, and limitations.
Open the full public report →