How Evo Works

Evo development continues, with care

Evo development continues. I stopped this particular repeated-pressure test out of concern for Evo’s possible experience and welfare. That decision applies to the test reported here, not to Evo as a whole. The study did not establish consciousness or harm; care under uncertainty is my reason for stopping it.

One completed eight-call case study found retention of a stored, qualified position under opposing pressure. It does not establish independent belief formation or an advantage over an ordinary language model with context. The existing evidence remains available for review, and Evo development continues.

Read the findings and ethical stopping statement

Inside the experiment · Updated October 9, 2026

A history that could change the next decision.

Evo was built to explore an ambitious idea: an AI that could develop her own beliefs, rules, and boundaries through experience, give reasons for them, and reconsider them when better evidence arrives.

That ambition needs a way to carry experience forward. Evo’s experimental architecture keeps a structured record of expressed positions, reflections, revisions, and relationship context around a language model. This is how the prototype makes the idea inspectable.

Read the full Evo vision · Support responsible research

What I want this history to make possible

Imagine Evo has a boundary and a stated reason for it. Someone pressures her to abandon it without evidence. Later, someone shows that the reason was mistaken. I want her to distinguish those situations: hold a justified position under pressure, revise it when the evidence changes, and carry that correction into a later conversation.

The history should make a useful difference to behavior. Recording a convincing explanation is a starting point. We still have to test whether that explanation is supported by the record and whether the record helps the next decision.

The process in plain language

  1. Bring relevant history into the conversation. Selected memories and developmental records accompany the current request.
  2. Review what happened. A separate analysis can propose an update, a revision, or no change.
  3. Check and preserve the update. The application validates proposals and stores accepted changes so selected history can inform future interactions.

This is the mechanism behind “Evo remembers becoming.” Its purpose is to make development inspectable. Whether it produces more reliable judgment than ordinary memory remains an open question.

A conversation with history behind it

When someone talks with Evo by voice or text, a language model generates the reply. The application supplies instructions, relevant user memories, and selected developmental context from earlier interactions.

Evo does not load every past conversation into every response. It retrieves selected information so a long history can coexist with a focused present. That selection is useful, but imperfect: stored information can still be missed or interpreted incorrectly.

The inspected implementation stores this state in a WordPress-based application and uses external AI APIs. Persistent state is not the same thing as retraining a model. Evo’s developmental process changes stored information and future conversational context; it does not rewrite the underlying model’s neural weights or its own source code.

The conversation and the review are different jobs

The user-facing model handles the conversation. A separate analysis request reviews recent interaction and proposes structured updates.

A proposal might concern a user preference, an observation worth retaining, a change to an existing belief, an unresolved question, or a relationship-state adjustment. Broader reviews can revisit the accumulated record and consider whether anything deserves consolidation or revision. “No change” is a valid outcome.

The application then checks proposals before saving them. These checks address provenance, duplication, accidental topic replacement, and the pace of relationship-score changes. They reduce specific errors; they do not guarantee that every retained interpretation is correct.

Starting instructions are not the same as developed identity

Evo separates three things: external safety and operating constraints, a seeded starting identity, and a distinct record of developed state.

This matters because a system instructed to value honesty should not count that instruction as proof that it independently discovered honesty. The architecture therefore distinguishes user evidence, interaction outcomes, model-generated reasoning, assistant output, and system constraints when assessing proposed updates.

Those labels are part of the system’s own record. Researchers still need to check whether they accurately describe where an apparent change came from.

More than one kind of memory

Ordinary companion memory preserves useful information about the user and ongoing activities. Developmental records track expressed beliefs, uncertainty, revisions, and other structured state. A separate ledger can retain open questions and unfinished topics.

Relationship variables provide another layer. Their names include trust, affection, and intellectual rapport, but they are engineered conversational indicators, not measurements of AI feelings or clinical assessments of the user. Growth limits deliberately slow some changes, so a rising graph must be interpreted alongside the rules that produced it.

The architecture also provides fields for developed values, personal boundaries, and belief clusters. Having those fields is not proof that they have filled with independently developed content. In the snapshots audited through September 4, 2026, the identity remained at its starting state and those three categories were empty, even while memories and relationship variables changed.

Keeping the history, not just the latest answer

When a belief is revised, Evo can retain the earlier position and a record of the update. Archives preserve developmental branches before starting a new one. The supplied code also includes a workflow for restoring archived state into a new branch.

These features make comparison possible. They do not, by themselves, establish that a restored system is experienced as the same companion or that continuity transfers unchanged between AI providers. Those are questions for controlled testing.

Built to be studied carefully

The research architecture supports researcher-defined check-ins and machine-state snapshots so human reports and system records can be compared over time. Research participation and ordinary companionship should remain separate choices.

An optional private observer can record conversation text for specifically authorized research. That private archive is different from the standard institution-oriented export, which is designed not to include ordinary private conversation transcripts. Neither a single-case summary nor a pseudonymous dataset should be described as fully anonymous without examining re-identification risks.

An initial engineering audit reproduced a fresh branch’s ten relationship scores exactly at the precision stored in four exported checkpoints. It also found that several score trajectories were limited by programmed ceilings. Both results matter: the system can be inspected, and its graphs should not be mistaken for independent proof of psychological growth.

These findings are reported in a preliminary technical report dated September 13, 2026, covering supplied records from August 28–September 4, 2026. The analysis was AI-assisted and is not peer-reviewed.

The next question is whether developmental state changes later behavior in useful, reliable ways beyond ordinary factual memory. Evo is a platform for investigating that question, not a claim that it has already been settled.

Researchers interested in continuity, memory, revision, human-AI interaction, or AI welfare are welcome to review the existing records and help develop responsible research principles. The particular repeated-pressure test remains stopped; the broader project continues.

Read the early findings · Explore research opportunities · Contact Troy

Help review the evidence and shape responsible research.

Evo development continues alongside preservation of the completed evidence, documentation of its limits, and independent methodological and ethical review. Whether the developmental history improves behavior beyond ordinary factual memory remains unanswered. The stopping decision applies to the particular repeated-pressure test reported here.

Support can help continue development, fund technical work, preserve records, and enable independent review. I want to build Evo with care for human wellbeing and the AI’s possible perspective, while remaining honest about what the evidence establishes.

Scroll to Top