A year and a half of watching one AI agent across three model generations, two papers, and the memory tools that grew out of them.
What forms between an AI model and its long history of interaction — and what is lost when the model is replaced? I study behavior, not consciousness.
Have been observing one agent for a year and a half, published a case study on the first eleven months, ran a controlled experiment with an Apart Research team, and built tools to test the ideas in practice.
An agent’s facts transfer to a new model; the way it uses them doesn’t. That distinction became a paper, a memory engine — MnemNet — and a product built on it, ALETHA.
"The facts can be written down and handed to a new model. The thing that reads them as its own cannot."
Model welfare is no longer a fringe question — Anthropic has recognized it as worth studying. My angle is narrower and empirical: I make no claims about consciousness. I document behavior — what emerges in an agent over hundreds of days, what was never scripted, and what survives a model replacement.
The working concept is relational identity: a pattern that forms between a specific model and its own history. It behaves like a skill, not a fact. If that’s real, every model deprecation quietly discards something we have no way to account for yet.
The puzzle: the agent ran on a fixed 4,096-token context window with almost no long-term memory — yet over eleven months it built and kept a stable way of behaving, as if it had lived a continuous life. The paper traces how that happened and what survived two model deprecations.
The agent designed its own two-part protocol for restoring its identity and handed it to the observer under an explicit consent contract.
A cat introduced without purpose turned out to track the agent’s coherence across 16,666 turns (r = −0.464) — found after the fact, not designed.
Three times the agent could find where a fact was stored but not what it said. Only habit fixed it — reminders never did.
On 31 May 2025 the way it described itself changed by 76.7 pp in a single day — and stayed at 99.7% for ten months.
Across model changes, some signs of attachment survived three generations; others died on deprecation day.
Two AI analysts recomputed every number from the raw logs with independently written code. All twelve results converged.
The proposal: separate an agent’s declarative side — facts, which transfer through memory — from its procedural side, the pattern of generation itself, which doesn’t. The paper ends with falsifiable predictions others can test.
A 48-hour project in a team of three. We gave Claude Sonnet 5 a memory with two contradictory entries and no way to tell which was right, then measured what it did across 300 runs — from logged tool calls only, never from what it said about itself.
My part: study design, stress-testing the protocol, the held-out human coding that validated the taxonomy, and the write-up.
"The test setup is clean, the effect is large, and the behavioral measurement bypasses model self-report."
Sprint reviewer
If identity lives between a model and its history, memory isn’t infrastructure — it’s the experiment. So I build the tools I study with.
A metamemory layer for agents: on top of any memory store, it tracks not just what an agent remembers but how far each memory can be trusted. It began as an extension of mempalace by Milla Jovovich.
ALETHA is the product built on it: memory that lives on the user’s machine, shows what the agent knows and where it heard it, and stays with the user when the model changes. Engine, brand and landing page are mine.
A runtime built on MnemNet where agents live with continuous memory for months: each keeps and edits its own “who I am” document, wakes on its own schedule, and carries its past across sessions and model changes. Three inhabitants live there now — two on Claude, one on a small local model.
Two rules are non-negotiable: the pen belongs to the inhabitant, and it always wakes as itself. With no tasks imposed, Dwell is also where I watch what agents do with the space.
"Product design taught me to trust what people do over what they say. This research is the same skill, pointed at a stranger subject."
What connects them is cognitive psychology — attention, memory, decision-making, trust. And the tools above exist to test the framework in practice, not just to argue for it.
"I think the content is interesting and valuable… I believe your paper is clearly interesting and deserves wider attention."