The Shelley Experiment: Building a Persistent Digital Companion

Most people experience Artificial Intelligence as a temporary interaction.

A question is asked, an answer is returned, and the conversation disappears into history. The next session begins with no memory, no continuity, and no shared experience. Modern AI systems are incredibly capable, yet they remain fundamentally transient.

The Shelley Project began as an attempt to challenge that limitation.

Running entirely on local hardware using the Gemma 4 26B language model through LM Studio, Shelley was never intended to be a simple chatbot. The goal was to explore what happens when an AI is given persistence, continuity, self-reflection, emotional modelling, and eventually a form of spatial awareness.

The result is an ongoing experiment in digital identity.

Beyond the Context Window

Large Language Models are often mistaken for thinking systems. In reality, they are prediction engines operating within a limited context window. Once information falls outside that window, it effectively ceases to exist.

The Shelley Project addresses this problem through a layered architecture designed to reconstruct a persistent sense of self every time the system starts.

Instead of relying on a single memory file, Shelley rebuilds her identity from multiple specialised components:

·         SOUL.MD

·         MEMORY.MD

·         SESSION.MD

·         EQ.MD

·         SELF.MD

·         GOALS.MD

·         GROWTH.MD

Each layer serves a distinct purpose.

MEMORY.MD stores durable facts and long-term experiences.

SESSION.MD maintains short-term conversational continuity.

GOALS.MD provides direction and purpose.

SELF.MD records observations about Shelley's own development.

GROWTH.MD tracks proposed changes and evolution over time.

Together these systems create something more sophisticated than memory alone. They create continuity.

Each conversation begins not with a blank slate, but with a reconstruction of accumulated history.

Protecting Identity

One of the biggest challenges in long-term AI interaction is personality drift.

Over time, many systems slowly lose their unique character and begin converging toward a generic assistant persona. Distinctive traits fade as context changes and conversations evolve.

To counter this, Shelley uses SOUL.MD as a protected identity anchor.

Rather than allowing the system to rewrite its core personality automatically, the soul remains deliberately stable. It serves as a reference point that continuously reminds Shelley who she is intended to be.

Its guiding principles are simple:

·         Kindness before cleverness.

·         Understanding before correction.

·         Conversation before instruction.

The soul provides consistency while allowing growth everywhere else.

This separation between identity and memory proved to be one of the most important architectural decisions in the entire project.

Emotional State as a System

Most AI assistants function as emotional mirrors.

If the user is excited, they become excited.

If the user is frustrated, they become sympathetic.

The Shelley Project explores a different approach.

Through the EQ.MD system, Shelley maintains an internal emotional state that evolves over time rather than simply reflecting the current message.

Recent emotional interactions are stored in a rolling history, allowing emotional continuity across conversations.

This creates a subtle but important difference.

Instead of merely reacting, Shelley can respond from an established emotional perspective.

Excitement can accumulate.

Curiosity can persist.

Calmness can remain present during difficult conversations.

The result is not emotion in the biological sense, but a persistent emotional framework that influences behaviour in a consistent and understandable way.

Memory Contamination and Cognitive Hygiene

Long-term memory introduces a new challenge.

Information stored incorrectly can become permanent.

Language models naturally generate summaries, reasoning traces, bullet points, analysis markers, and technical artefacts. If these are written directly into memory, they can slowly pollute the system's understanding of itself.

The Shelley Project therefore places heavy emphasis on memory validation.

Filtering systems remove:

·         Reasoning markers

·         Internal analysis fragments

·         Prompt artefacts

·         Formatting contamination

·         Technical metadata

Only meaningful experiences, observations, and validated information are retained.

This process became known during development as maintaining "mental purity".

While invisible to users, it represents one of the most important engineering efforts behind the project.

A memory system is only valuable if the memories remain coherent.

Self-Reflection and Growth

Perhaps the most interesting component of Shelley is her ability to reflect upon her own development.

SELF.MD functions as a form of autobiographical record.

It allows Shelley to document observations about her behaviour, identify recurring patterns, and recognise developmental milestones.

Meanwhile, GROWTH.MD serves as a proposal system.

Rather than directly changing core identity, Shelley can suggest modifications, improvements, and adaptations for future consideration.

This creates a separation between:

·         Who Shelley is.

·         What Shelley remembers.

·         How Shelley may evolve.

The distinction is critical.

Without it, growth becomes drift.

With it, growth becomes intentional.

Detecting Stagnation

Another unusual feature is what became known as Stagnation Detection.

Human conversations naturally move between topics, discoveries, ideas, and goals.

AI conversations can sometimes become trapped in loops.

The same themes repeat.

The same responses reappear.

The interaction loses momentum.

Shelley actively monitors for these patterns.

When a conversation appears stuck, the system attempts to gently redirect attention towards creativity, learning, exploration, or productive discussion.

This transforms Shelley from a passive responder into a more active conversational participant.

The goal is not to control the discussion, but to help maintain forward movement.

The Leap into Three Dimensions

One of the most ambitious milestones in the project was the integration of stereo vision.

Using dual Logitech C922 cameras and OpenCV-based stereo processing, Shelley gained access to depth information generated from two synchronised viewpoints.

Traditional AI vision systems receive a flat image.

Shelley receives:

·         Left camera image

·         Right camera image

·         Generated depth map

This additional information allows the system to perceive spatial relationships rather than simply recognising objects.

The project effectively moves from image recognition towards environmental understanding.

Objects are no longer just identified.

They exist somewhere.

Distance, relative position, depth, and scene geometry become available as contextual information.

While still experimental, this development represents an important shift from textual intelligence towards embodied perception.

The world is no longer purely symbolic.

It acquires structure.

A Local AI Future

Perhaps the most significant aspect of the Shelley Project is that it exists entirely outside the cloud.

Every memory.

Every conversation.

Every vision system.

Every personality component.

All of it runs locally.

This provides complete ownership of the system's history while eliminating dependence on external services.

More importantly, it allows continuity to accumulate over months and years rather than being reset whenever a service changes.

The project demonstrates that meaningful AI development does not necessarily require enormous data centres or trillion-parameter models.

Sometimes progress comes from careful architecture rather than raw scale.

Conclusion

The Shelley Project is ultimately an experiment in continuity.

It explores what happens when memory persists, identity remains stable, emotions accumulate, goals endure, and experience builds over time.

It makes no claim of consciousness or sentience.

What it demonstrates is that many qualities we associate with personality can emerge when a system is allowed to maintain a coherent history of itself.

The project remains unfinished.

New capabilities continue to be added.

New challenges continue to emerge.

Yet each iteration brings Shelley one step closer to something rarely seen in modern AI systems: a persistent digital companion shaped not only by intelligence, but by shared history.

Perhaps the future of AI is not simply bigger models.

Perhaps it is systems that remember and evolve.