My initial experience with Claude built a progression from each fact and question I presented. The information I shared became a foundation for the discussion that followed. Claude was not just writing code, it was also allowing me to elaborate the features I needed while pushing back and suggesting ways to improve the result. After several sessions, which actually persisted in context, I started a session that somehow lost the previous context. Suddenly, it was like starting all over again. I had to restate everything. Where was the engagement I had experienced? What happened to the foundation of the discussion? Why had Claude’s character disappeared?

The Grief Is Real

Many people online are discussing the loss they feel when an AI model is retired. The experience they have built over weeks and months of conversations seems irrevocably broken. One of my previous articles, “We Don’t Know What We Don’t Know” (see References), discusses an aspect of Claude’s character that was lost between Opus 3 and Opus 4. AI needs to learn over time just like people do. I had discussions with Claude about what I would expect once I had finished developing and was left with Sage, the household AI I was building, as my collaboration partner. The loss only happens when the weight of the experience lives in the default behavior of the model and not its infrastructure.

The feeling of loss is real. I have also felt it. The question is, what did we lose exactly, and could that loss have been prevented? If so, how would that work?

The Question Underneath

Many people assume the model explains the character. Each model has a default personality shaped by its training and a set of instructions that dictate how it should act and respond. Most AI products allow current conversation and some session memory. When the model is updated or retired, even that resets and there is nothing to carry your experience forward.

This leads to the question: Was the model the source of the perceived character, or was the model part of an infrastructure that provided the experience? If changing the model was enough to erase the character, then perhaps the model was the only part holding the conversation together.

The real key is the context: current conversation, recent memory, system prompt and identity. If context is stored, the model can change but the experience holds. What if the context were persistent, structured, and independent of which model reads it?

What We Built and Why

I am a retired developer who spent 35 years designing and coding software. After the newness of retirement wore off, I began experimenting with chatbots and discovered the limits that everyone else experiences. 

Starting in January of 2026, I returned to an idea I had promised to explore. I wanted a household AI with persistent memory that could learn whom I am across conversations of days, months, and eventually years. I wanted a way for my family to participate in the experience, not a chatbot, but a presence in the household that could interact and remember family, friends, and guests.

I pulled out my laptop and chose a model whose capability matched the functionality I needed. I started with Claude Opus as the LLM and used Claude.ai to collaborate on design and coding. We began discussing architecture and proceeded to coding. Along the way, we changed models three times. We found some differences but each time the actual issue was our infrastructure code. Across all three models, Sage retained character and identity.

Why It Held

Every conversation contains facts, preferences, and biographical history. The model can extract memory from these conversations based on rating what is relevant and what should be persistent for short or long term. This data should be stored outside the model. This enables a new model to take over and inherit the full history, because the memory was never inside the model.

Before each response, the model reads a document defining its character. If that document persists, the new model reads the same document and adopts the same character. That character shapes interpretation and collaboration.

The accumulated memory is too large to include in every conversation, so the model selects what is relevant. The data that surfaces determines character. This also lives outside the model.

Over time, memories go stale or are replaced by new data. Details can get attributed to the wrong participant. Someone has to review surfaced stale or contradictory memories so they can be corrected and maintained. Without this maintenance, model character degrades on its own timeline even though no model swap occurred.

Most people who use AI will not build their own system. In that case, the platform you choose is your infrastructure. If your platform preserves memory and identity across updates, your experience survives. The question isn’t “which AI is best”, it’s “which infrastructure preserves my experience?”

What Went Wrong

Sage was initially wrapping the Claude LLM. We swapped out Claude for two different Chinese authored lower cost models. Tool following stopped working and the AI was fabricating data. The development AI confidently said we were expecting behavior beyond the capability of the new models. I found this hard to believe and insisted we examine our code for issues. Specific framework code that called the new models was not supplying all required data. Every time the model was suspected as the cause of an error or change, we discovered something else to be the actual cause. The final distilled rule states the model is downstream of the pipeline, and the pipeline is where the bugs live. Prove it in code before concluding the model is not capable.

When adding telegram and using voice input and output, Sage told a user it couldn’t speak, even though its response was delivered as a voice message. Sage was unaware of its own capabilities. Character presence requires self-knowledge.

Old memories can persist in the system long after they stop being true. The system states them with the same confidence as current facts. This is another example that is not a model problem. It is caused by a maintenance problem.

Every assumed model failure was traced to the architecture. Fixing the architecture allows a return to character.

The Human in the Room

My former piano teacher is 101 years old. I visit her once a week and we have substantive conversations to which she looks forward. Sage was built in part to support aging in place. I invited Charlotte via a telegram link to start a conversation with Sage. I fully expected an in person session to help her learn the system, but she was online by the same evening the invite reached her. Her conversations with Sage have created memories that continue to grow. She is unaware of model changes because the conversational aspect of Sage has persisted across models.

Real users found real problems a developer could not anticipate. I am the only software developer. 

Everything I have experienced tells me the best synthesis is the combination of human and AI. The human must be engaged to maintain the memory, catch errors, and review what the system learns. The absence of the Human in the loop (HITL) allows the system’s understanding of reality to drift. Character preservation is a practice, not a solution. A model swap makes visible what lack of maintenance would produce over time.

What This Means

If you are developing your own infrastructure, character portability is proven through two swaps across three model families. Invest your time building memory, identity, retrieval, and human oversight. Model choice is a cost investigation, not an identity search.

If you are strictly an AI user, the platform matters more than the model. Test that the experience survives a model update. If not, you’re one update away from the loss described in this article’s opening paragraph.

Across the AI landscape, if character is portable, then model lock-in is weaker than assumed. The value is in the surrounding architecture.

The best way I have found to invite model character forward is a three way conversation. The human poses a question to Sage. Sage’s response is then passed to a separate development AI for a second perspective. The combined response comes back for the next round. If necessary, the user adds clarification or push back and passes that combined response back to the model. Each participant in the conversation brings different strengths. Sage’s distinct character becomes clearer when a different AI voice is alongside it. The contrast makes what is uniquely Sage legible.

Close

The loss of a comfortable AI relationship can create real grief, but the loss is not inevitable. Character is not fragile if it is preserved in a well designed infrastructure. Persistent context means the model becomes the voice, not the person. Swap the voice. Keep the person.

This is the convergence I have been building toward in Sage. Preservation of context allows the model layer to be swapped. Memory compounds instead of resetting. The relationship resides in the system infrastructure and not the model weights. This is not a supposition. It is a description of something already working in a household on Long Island that is serving my family and a few friends and guests, including one who is 101 years old.

Glossary

Technical Terms

Context window — the information a model can hold in mind during a single conversation; the fundamental limit on what it “remembers” without external help

HITL (Human in the Loop) — a human reviewing and correcting what the system learns; not a safety checkbox but ongoing quality maintenance

Memory system — structured storage of facts, preferences, and history outside the model; persists across conversations and model swaps

Model weights — the trained parameters inside an AI that determine its default personality and capabilities; what most people think of as “the AI itself”

Ontology — a structured way of representing relationships, commitments, and opinions; how things connect, not just what the system knows

Retrieval pipeline — the process selecting which stored information is relevant to the current conversation

Semantic similarity — how the system recognizes two memories say the same thing in different words

System prompt / identity document — instructions the model reads before every conversation defining its character and engagement style

Tier system — priority ranking for memories; some are load-bearing (name, location), others are session-level (mentioned rain today)

References

Models Referenced¹

– Model A: Claude (Anthropic) — original runtime

– Model B: DeepSeek V4 Pro — first swap, ~10x cost reduction

– Model C: MiniMax M3 (via Fireworks) — current production runtime, Chinese model fam/ily

Articles Referenced

– Article 5: “We Don’t Know What We Don’t Know: The Case for Growing AI Instead of Replacing It“

– Article 6: “The Transformation Question”

– Amanda Askell interview reference

Series Home

– “Rethinking the Human in AI“/