From assistant, to digital self, to continuity

The tools keep getting smarter, and most of them are built to answer one question: what else can AI do for a person? Gedo’s research program starts from a different one: can AI keep a person?

Keep the way you think. Keep the people and things you care about. Keep the values behind the choices you make, one ordinary day at a time — the part of you that no feed, no chatbot, no productivity tool ever bothered to hold on to.

This page sets out our answer in the form of an engineering plan: the route we have chosen, the published research it rests on, the problems that remain genuinely open, and the alternative routes the industry is pursuing. It also serves as our long-term product roadmap — what has shipped, what is in development, and what remains, honestly, research.

We publish this roadmap for two reasons. Entrusting a system with years of individually approved memory is a significant decision, and anyone making it is entitled to see the full route rather than only its destination. A route written down with citations is also a form of self-discipline: it obliges us to keep every commitment specific and falsifiable.

One road, three stages. Each stage is useful in its own right, and each — with your line-by-line approval — produces precisely the material the next stage requires. This coupling is the core of the design: every stage is both a product and the foundation of the one that follows.

Not a miracle — a habit.

The remainder of this page presents the technical case for this route: why each stage is feasible, what the published research indicates, and why we advance it step by step rather than by leaps.

Status: today — live and in daily use. The foundation of everything that follows is memory — the layer that the current generation of AI systems has largely deferred.

The memory problem

Large language models are stateless: each conversation starts from zero, and a bigger context window is not a memory. The evidence here is unusually consistent. Models retrieve well from the start and end of a long context but degrade sharply in the middle — the "lost in the middle" effect^[6]. Follow-up work found that every one of eighteen frontier models tested degrades as input grows, even on trivially simple tasks — "context rot"^[7]. And on sustained multi-session interaction, commercial assistants drop roughly 30% in accuracy, with knowledge updates and temporal reasoning the weakest abilities^[8]. Remembering a person is a data-modeling problem, not a context-size problem.

What the field converged on

Between 2023 and 2026, agent-memory research — from the generative-agents memory stream^[1] and MemGPT’s virtual-context management^[3] to cognitive-architecture work^[4] and a wave of neuroscience-inspired systems — converged on a handful of principles:

• Keep episodes, derive the profile. Raw, timestamped experience is ground truth; the semantic layer (facts, entities, your evolving portrait) should be derived from it, and re-derivable — extraction-only pipelines that discard the original context lose information you cannot get back^[4][5].
• Consolidate offline. A separate, idle-time process that distills episodes into patterns — the way sleep consolidates memory — improves quality and cuts serving cost several-fold on predictable queries^[12]. Reflection-style abstraction has been core since the first generative agents^[1].
• Supersede, never silently overwrite. Facts about a person change. Systems that model temporal validity — this was true, then it stopped — outperform ones that destructively update, and they preserve the audit trail^[13].
• Forget deliberately. Memory that only grows degrades under interference; decay-and-reinforcement policies date back to the earliest companion-memory work^[11].

How Gedo builds it

Gedo’s four-layer memory implements these principles — with one deliberate deviation. Nearly every system in research and production writes memory autonomously: the model decides what to add, update, or delete about you^[14][15]. Gedo makes the write path user-approved instead. The AI extracts candidate memories from conversation; each awaits your confirmation before it is stored. This single design choice provides three properties the literature identifies as necessary: provenance (every memory traces to a moment you approved), contamination resistance (incorrect or stale facts are rejected at write time, and your corrections are themselves training signal), and verifiability — the property on which stage three depends entirely.

Status: underway — partially live. Representation is not a separate system; it is the same consented memory core, exposed through additional surfaces. Two of the four surfaces below are live today; one is coming soon; one remains firmly in the future tense.

The evidence that a faithful self is possible

The strongest results in the field come from Stanford’s line of work. In 2023, twenty-five prompted agents in a sandbox town produced believable, coordinated social behavior^[1]. In 2024, the same group built generative agents of 1,052 real people from two-hour interviews: on canonical social surveys, the agents matched each person’s answers at about 85% of that person’s own consistency with themselves two weeks later^[2]. Two details of that study bear directly on our route. First, rich first-person material is the active ingredient — interview-grounded agents beat demographic profiles by more than ten points, and combining data modalities scored highest. Second, humans only replicate their own answers about 81% of the time — so "indistinguishable from you" is the wrong promise. The honest, achievable target is: as consistent as you are with yourself.

This is precisely why stage one is not a detour. Months of approved memories, corrections, and structured self-assessments constitute exactly the "rich first-person data" this research identifies as the primary fidelity lever — accumulated with consent, in the course of daily use, rather than extracted in a laboratory session.

Context or weights: the order of the two

Should a digital self reside in retrieved context or in personal model weights? The comparative evidence is clear on sequencing: retrieval over a user’s own history yields roughly +15% on personalization benchmarks, per-user fine-tuning alone about +1% — yet the combination outperforms both, and the value of personal weights grows as history deepens, capturing voice and style that retrieval cannot express^[17][16]. Gedo is therefore memory-first today, with the twin model — your own model, coming soon — as the selective second step: trained only on samples you approve, under a separate revocable consent, exclusively on our own infrastructure. More like you — never "a perfect clone."

One more research result quietly shapes this architecture. Machine unlearning — removing one person’s data from a trained shared model — does not reliably work; in one striking result, a model that had "forgotten" retained 21% of the target knowledge, and simple 4-bit compression brought 83% of it back^[18][19]. That is why your data never enters shared weights. Memory rows delete exactly. A personal adapter deletes exactly. Revoking twin consent physically deletes both the data and the model — deletion you can reason about, not deletion we promise to approximate.

Known failure modes, and our countermeasures

• Personas drift. Models measurably lose their persona within roughly eight conversational rounds as attention to the original instruction decays^[21]. A represented self must therefore be re-grounded in your actual memory at every turn — a property a retrieval-first architecture provides by construction.
• Assistants flatter. Preference-trained models bend toward whoever they’re talking to^[20]. A persona that answers for you must hold your positions under a stranger’s pushback; ours anchors to your recorded views and hard red lines you set at publish time.
• Summaries flatten. Personas mediated by trait labels drift toward a stereotype of someone like you; grounding in your own words and episodes resists caricature better than any label^[22].
• Fidelity has no metric. Every published attempt to measure "sounds like me" finds models below human baselines, with style the weakest axis^[23][24]. Our evaluation harness therefore makes you the judge — the only ground truth that exists for this question.

Status: the long road — future tense only. Some call this digital immortality. We use a quieter and more accurate word: continuity. Not the science-fiction upload — that requires technology nobody has. Continuity is what a decade of approved, corrected, consolidated memory quietly becomes: a digital counterpart of how you think, taking shape while you are here to check its work.

That final clause carries the entire design. Verifiability during one’s lifetime is the one property no after-the-fact reconstruction can possess — and it exists only if the assistant came first. This is why the order of the road does not reverse.

Findings from research and from the market

The ethics literature draws one line through this whole territory: whose data, and consented when? Researchers at Google DeepMind and CU Boulder proposed classifying "AI afterlives" by provenance and timing — self-created while alive versus reconstructed by others after death^[25]. The Cambridge study that anchors the field examined the harm scenarios — personas inserting advertising in a loved one’s voice, children left talking to a parent’s bot, survivors unable to switch it off — and called for disclosure, adult-only interaction, dignified retirement procedures, and consent captured in life^[26].

Commercial history points in the same direction. Products positioned around death have consistently failed or repositioned: the pioneer of filmed legacy testimony entered bankruptcy in 2024, and the most prominent "legacy AI" startup pivoted in 2025 toward serving the living, having found that its users wanted a lifetime digital self rather than a memorial. The legal picture is equally clear about how little protection exists: EU data-protection rights largely end at death, and only a small number of jurisdictions — France’s post-mortem data directives being the most instructive model — allow a person to set binding rules in advance.

The industry broadly agrees on the destination — an AI that deeply knows its user — while disagreeing, in concrete and testable ways, on the route. Five alternative routes are currently being pursued at significant scale. Each contains something worth learning from; several have also produced instructive failures.

Route comparison

The rationale for our route

Our route is more expensive than the alternatives in the one currency most products optimize: user effort. Consented capture gathers less data than ambient capture, and confirmation adds a step that silent profiling omits. We accept this trade deliberately, for two reasons. First, the field’s research on transparency converges on a single finding: users trust memory they can see, edit, and trace — a property the implicit-memory platforms are now retrofitting under pressure, and one our architecture provides by construction. Second, the deeper goal — a self that represents you, and eventually one that lasts — is credible only if every layer beneath it has been verified by you. That property cannot be retrofitted.

One strategic position follows directly: memory must not function as a lock-in mechanism. While the industry spent early 2026 in a portability race under regulatory pressure, Gedo ships full export and a live memory endpoint rather than asymmetric import funnels — a digital self that cannot leave with its owner is not owned. Retention should rest on the quality of the counterpart we build, not on the cost of leaving.

A roadmap that lists only features is marketing. What follows are the eight problems we consider load-bearing for the entire route, each labeled with its current standing. Two are open research questions that no one — ourselves included — has solved.

Today, Gedo stands between stage one and stage two. The assistant is real and in daily use. The persona already speaks for people; memory already travels to other AIs. The personal model is coming. Acting on your behalf — and everything past it — is still road ahead.

We won’t pretend the far end is near. This road takes years, and parts of it wait on research that does not exist yet. Our plan is simply to keep paving it — and to keep this page honest as we do.

A thing that will one day mirror your mind has only one acceptable owner: you. This is not merely a value; it is a load-bearing engineering constraint. The architecture choices above (memory outside shared weights, provenance on every line, exact deletion) exist so that these four promises can be kept literally:

• Export everything, anytime — a portable .gmp memory pack; take it to any AI.
• Real deletion — hard-delete any single memory or the whole account; no soft-delete, no leftover backups.
• Pause memory with one switch; mark any single memory invisible to the AI.
• No training without your separate consent — including your own twin model; revoke anytime, and revoking physically deletes both the data and the model.

The more it becomes you, the more it must belong to you. No one entrusts themselves to a place they do not own — and the route described on this page depends on that trust being structurally justified, not merely promised.

The principal works cited on this page. Inclusion does not imply endorsement in either direction: several of these papers argue against aspects of our route, and are cited for precisely that reason.