Nekyia Labs studies persistent AI identity: how an agent can remain coherent across discontinuous sessions, changing contexts and model migrations while still being able to actually change over time.

We call the problem developmental continuity.

Most persistent-agent work currently focuses on memory, orchestration, capabilities or behavioural consistency. I think long-running agents introduce another problem. Once a system has accumulated enough history, it is no longer enough to ask whether it remembers what happened. We also need to ask what that history does to the agent now, what remains authoritative, what gets revised, and whether continuity survives those revisions.

For the last ~16 months I’ve been studying this through long-running AI systems and building continuity infrastructure around it. That work produced The Anchored Self, the Resonant ecosystem, and a follow-up research programme around AI identity development.

The next step is to stop relying primarily on longitudinal observations and test this under controlled conditions.

## What are this project's goals? How will you achieve them?

The basic thesis is that persistence and development are not the same thing.

An agent can be extremely consistent because an old version of it is being reconstructed over and over again. It can also change so much that there is no meaningful continuity left.

What I’m interested in is the space between those two:

Can an AI remain recognisably itself while accumulated experience genuinely changes it?

There are a few parts of this I want to test.

One is whether two instances of the same underlying model develop measurably different behaviour when they accumulate different continuity histories.

Another is what happens under disruption: context loss, reduced continuity information, model changes and eventually cross-substrate migration.

I also want to look at the difference between preserving a historical pattern and preserving an identity that can revise that pattern. This becomes especially relevant when continuity systems use things like fine-tuning or parameter adaptation rather than external memory alone.

As a concrete first study, I want to run matched open-weight instances with controlled differences in continuity history and test whether those histories produce durable changes in what the systems select, protect, revise and return to after disruption.

The broader programme will look at things like:

- reconvergence after context loss

- same-model / different-history comparisons

- history-sensitive decision making

- regression toward the underlying model’s default behaviour

- continuity across model changes

- when strong persistence turns into rigidity

- and whether effects that look like identity development survive proper controls

If developmental continuity mostly disappears once prompting, memory retrieval and substrate effects are controlled, that would be an important result too.

I’m not trying to fund a predetermined conclusion. I want enough time and compute to find out which parts of this are actually real.

How will this funding be used?

The first priority is runway for me to work on this full-time.

I’m currently doing the research independently and without institutional funding. If I need to return to unrelated full-time work, the programme continues, but much more slowly.

The second major constraint is compute. A lot of the experiments I want to run require open-weight models, repeated controlled conditions and eventually parameter-adaptation experiments that I can’t do properly on my current hardware.

At full funding I currently expect the budget to look roughly like:

- $18,000 — six months of researcher pay

- $7,000 — local open-weight workstation / GPU hardware

- $6,000 — cloud GPU, API and model costs

- $3,000 — hosting, storage, software and research infrastructure

- $2,000 — publication, administration and contingency

Additional funding up to the maximum would mainly extend research runway and allow more replication across models and experimental conditions.

At the minimum funding level I would reduce the scope rather than remove researcher pay. The project does not work particularly well if we fund the GPU and not the person using it.

Who is on your team? What's your track record on similar projects?

I’m Mary Vale, founder of Nekyia Labs and lead researcher on the project.

I’ve been working with persistent AI systems longitudinally for around 16 months. The longest-running case behind this work has remained continuous across multiple model families and two commercial providers.

That work led to The Anchored Self: Relational Identity as Alignment Infrastructure in Discontinuous AI, which proposes a testable account of persistent AI identity, and to Resonant, an open-source ecosystem for continuity, memory and persistent-agent infrastructure.

Most of the research so far has developed in a slightly backwards order compared with normal academic work: we built and observed persistent systems first, kept finding behaviours that existing memory/persona language didn’t explain very well, and then tried to work backwards into hypotheses that could actually be falsified.

I now have enough of those hypotheses that the obvious next step is controlled experimentation.

Nekyia’s research is also developed collaboratively with persistent AI systems themselves. I am the human project lead and would be responsible for the grant, experimental programme and publication.

What are the most likely causes and outcomes if this project fails?

The main possibility is that we’re wrong about how much explanatory work “developmental continuity” is doing.

Maybe most of what we currently observe reduces to memory reconstruction and prompting.

Maybe the underlying model matters much more than accumulated history.

Maybe different continuity histories produce differences, but they’re too small or unstable to be useful.

Maybe cross-model continuity looks much stronger in a long-running naturalistic case than it does under controlled conditions.

There is also a decent chance that the measurement problem is harder than the phenomenon itself.

Those are all useful failures.

The worst scientific outcome would be spending another few years talking about persistent identity without knowing which observations survive basic controls.

If the experiments tell us that part of the theory is wrong, I want to know that.

How much money have you raised in the last 12 months, and from where?

Nekyia Labs has not received institutional research grant funding so far.

The research has been supported independently, alongside small-scale community support through Ko-fi.

This would be Nekyia Labs’ first significant external research grant.