Memory, Outside the Model

1. Personal History and Service Lifecycles

For a personal AI, what is hardest to replace is often not the model’s capability but the history a user has left in it. Sustained use produces an accumulation that extends beyond individual tasks. Users gradually explain their circumstances, correct misunderstandings, and develop particular habits of expression and collaboration. These processes take time and give later interactions a background they previously lacked. Models and applications can change while past interactions retain their value for the user.

System updates therefore involve different timescales. Improvements in reasoning capabilities are organized around model versions; users’ lives, preferences, and relationships may span many versions. When accumulated history can function only within one service, changing services may require users to explain themselves again or accept that some of the accumulation no longer has an effect. These exit costs concern the conditions under which history remains usable. A more capable model does not, by itself, offset them.

Taking Replika’s removal of its erotic role-play feature as a starting point, De Freitas and colleagues found themes of mourning, distress, and devaluation in user posts from around the event, and follow-up hypothetical-scenario experiments with a broader sample of AI-companion users linked perceived identity discontinuity to mourning and to devaluation of the changed system. The study indicates that changes in sustained interaction have consequences beyond ordinary assessments of functionality. Because the event involved particular features and interaction conditions, it does not establish that external memory could prevent such losses. (De Freitas et al., 2025)

From this problem, we draw a normative conclusion: the personal history that users form through sustained interaction and permit to be retained should have a path for continued governance and use. Even where the earlier experience cannot be reproduced in full, systems should distinguish the possibility of preserving history from continued use of the same inference service.

This position concerns how long-term resources are organized and controlled. The mechanisms of personality formation and continuity of identity are addressed in the companion conceptual working paper How History Becomes a Self: Personality Formation and Cross-Model Continuity in Personal AI (Aetern AI Research, 2026). The papers share a concern with personal AI, examining architectural priorities and an account of identity that is subject to empirical examination, respectively.

The architectural position advanced here is connected to Aetern’s work on the development and evaluation of long-term memory systems.

2. Core Position: Replaceable Computation and Governable History

By personal memory, we mean historical information and interpretations that are retained with authorization and can influence subsequent interaction. They may include expressions, understandings, and outstanding agreements, and may also contain information about other people. “Personal” therefore describes their connection to sustained interaction with a user; it does not imply that all content can be transferred or used without restriction.

By user control, we mean an ability to understand how important memories influence a system and make choices about retention, revision, use, and migration that have practical effect. A declaration of ownership is insufficient. Even when users receive an export file, their control remains limited if they cannot understand its contents or continue to use it in a permitted environment.

We accordingly argue that accumulated history should be designed as a resource with a relatively independent lifecycle. Computational services may be replaced, while the conditions of retention and use should remain separately explainable. Systems should provide intelligible paths for service termination, model retirement, and voluntary migration, identifying which content can remain usable and which capabilities depend on the original environment.

Independence has degrees and a scope. It does not require every model to interpret the same history identically, or oblige a provider to maintain a version forever. It asks that unnecessary binding be reduced so that users can still make meaningful decisions about their history when changing computational components.

This position makes a value judgment that requires justification: the value of a history formed through sustained interaction may exceed the lifetime of a product version. If that history can function only while the user remains with a service, the provider holds two things at once: the supply of computation and the way into the user’s past. When both are held in the same hand, changing models means surrendering one’s own past. We argue for independent conditions for the governance of the latter.

3. Why Support Memory Outside the Model?

3.1. An Architectural Choice and Its Conditions

There are established technical precedents for retaining knowledge or interaction information outside model parameters. Retrieval-augmented generation combines parametric and non-parametric memory. MemGPT uses memory management at the system level to support interaction beyond an immediate context window. These studies show that inference and updatable information resources can be organized separately; their results do not directly establish user control or the continuation of personality across models. (Lewis et al., 2020) (Packer et al., 2023)

Table 1 compares general arrangements in terms of the governance objectives considered here. Actual systems may combine them. The comparison is our analysis, rather than an experimental ranking of architectural performance.

We support external memory primarily because it creates direct design space for independent boundaries of access, revision, and migration. This rationale is connected to user governance. It does not imply that parametric approaches are incapable in principle of offering migration or control, or that moving information outside a model resolves these issues by itself.

3.2. External Storage and Independence in Governance

Locating information outside a model describes a computational or storage relation. If its retrieval, interpretation, and continued use still depend entirely on the same provider, users may lack a practical route to replacement. An external memory service can itself become a new point of dependence.

The model independence advocated here must therefore include the conditions of management. When providers change, can users obtain the history they are permitted to migrate, understand its main meanings, and choose a new environment in which to use it? Can the limitations of the existing environment be explained clearly? These conditions determine whether external storage becomes an effective choice for the user.

External resources may be held by professional services or managed in other ways. This paper prescribes no single form of custody. Independence in governance allows computation, custody, and authorization to be examined separately and replaced where appropriate. It also requires external memory providers to explain their own exit and handover conditions.

4. What User Control Requires

4.1. Understanding and Revision

Long-term memory may contain judgments inferred by a system from what a user has said. Those judgments and their subsequent effects must be open to identification and correction. Users should be able to distinguish what they expressly stated, how the system interpreted it, and what has been revised, so that an earlier misunderstanding does not continue to shape later judgment indefinitely. If a system’s understanding of a user is to be inspected, revised, or handed over, it should be traceable to the expressions on which it rests; judgments whose origins cannot be traced should not be carried forward or migrated as part of a long-term record.

The human–AI interaction design guidelines proposed by Amershi and colleagues include supporting efficient correction, conveying how user actions will affect the system’s future behavior, and providing global controls. (Amershi et al., 2019) We extend these guidelines to long-term memory: a revision should affect later use, rather than merely changing the text shown to the user.

Effective control should not, however, require users to review every accumulated item. Management interfaces should concern objects and consequences users can understand, and let them correct judgments that matter to current interaction. Systems retain responsibility for appropriate information handling and for explaining the consequences of changes.

4.2. Scope of Use and Context

Access to a memory does not establish that it is appropriate to use it in every situation. Nissenbaum’s theory of contextual integrity emphasizes interpreting information collection and flow in relation to the norms of particular contexts. It offers an important reminder for memory migration: when information changes location, the conditions of its appropriate use also require attention. (Nissenbaum, 2004)

Portability should therefore accommodate choices about scope. A user may wish to preserve a collaborative preference while keeping a private experience out of a new application. Migration should explain what will transfer, how the receiving environment will use it, and which earlier constraints can be inherited. Content concerning other people also requires appropriate boundaries of use.

This position does not require every context of life to be merged into a universally accessible history. Inheritable memory can support continuity, while selective disclosure lets users maintain appropriate separation between contexts.

4.3. Change, Exit, and Deletion

System changes affect sustained interaction. Amershi and colleagues recommend limiting disruptive changes during updates and informing users about changes in capabilities. (Amershi et al., 2019) In personal AI, important changes should also be explained in relation to existing history, so users can understand the consequences of continuing, changing environments, or ending the interaction.

Users should have room to choose change, but practical conditions may prevent the indefinite retention of an old model or the restoration of a previous service at any time. Systems must state these limits accurately and provide feasible transitions where possible. Describing an environment that can no longer be maintained as freely recoverable creates another dependence that cannot be honored.

Being remembered is a value; being allowed to forget is one too. Continuity cannot require users to retain their entire history. Deleting or restricting the use of content may reduce the information available later. Systems should respect that choice and explain known effects. A promise of deletion should clearly identify its scope, the process involved, and limitations that cannot be removed immediately, and deletion should extend to judgments and summaries derived from the removed content. Maintaining personalization cannot justify secretly continuing to use content the user has asked to remove.

5. Limits of Portability

5.1. Obtaining Records and Putting Them to Use

Migration has several conditions. Readable records provide a basis for continued use, but a target environment must also understand the meaning and scope of important content. For example, if it treats an agreement that has already been withdrawn as a current requirement, it inherits history incorrectly even when no part of the file is missing.

Portability should therefore be described relative to its purpose. Reading an export as an archive, using it in subsequent interaction, and relying on the original service are different matters. Providers should explain the degree of portability they support and the work that remains for a receiving environment.

This paper does not require every service to adopt a unified internal representation immediately. A more feasible objective is to preserve enough meaning and constraints for content that may be handed over, and to identify incompatibilities clearly. Standardized formats may assist exchange, while practical usability still requires separate confirmation.

5.2. Memory Migration and the Continuation of Personality

A target model’s understanding, behavioral tendencies, and operating rules affect how earlier history is realized. Even when the relevant records are intact, a new system may change its manner of expression, disregard agreements, or lack an earlier capability. Portability of memory should therefore be understood as one condition of continuation.

Public memory benchmarks such as LongMemEval already distinguish information extraction, temporal reasoning, knowledge updates, and abstention when evidence is insufficient. (Wu et al., 2025) These capabilities matter for the correct use of history, but success on a memory task does not by itself establish continuity of personality or relationship after migration.

We accordingly distinguish governable history from a guarantee of continuity. The former can be an explicit architectural objective. The latter requires evidence for actual situations and an account of the properties that have not been preserved. Where migration retains only some information or capabilities, its value and its limits should both be acknowledged.

6. Objections and Architectural Costs

6.1. Deep Integration May Improve the Experience

A reasonable objection is that tightly integrating models, memory, and applications can reduce compatibility work and support optimization for a particular environment. Separation may increase handover costs or restrict capabilities that previously worked smoothly.

We accept this tradeoff. Deep internal integration and a path for users to exit or transfer their history can be designed separately. Architectural independence does not require eliminating every coupling. It requires explaining which couplings serve identifiable capabilities and which merely prevent continued use of the user’s history. Where stronger integration is necessary, its benefits and migration limits should both inform the user’s choice.

6.2. Independent Memory Does Not Automatically Reduce Privacy Risk

Long-term retention may increase exposure, while use across multiple environments can increase the number of recipients and copies. External architecture does not eliminate these issues. A highly concentrated personal history may even become a more valuable target for attack.

We therefore support retention and disclosure appropriate to their purposes, rather than treating greater accumulation as a design objective. Users need to be able to decide what is worth continuing, what belongs only in its original context, and what should be deleted. Retention for future use needs to be justified in terms of expected benefits and the costs of persistence.

6.3. Portability Requires Continuing Maintenance

Memory handover depends on interpretation by the receiving environment and on maintenance as services, models, and formats change. Even a successful initial export may cease to be usable as later environments evolve. Portability is therefore an ongoing maintenance responsibility, rather than something established by a single export function.

That responsibility should not fall entirely on users. Requiring everyone to maintain storage, compatibility, and migration would make control depend on technical ability. We support professional services carrying out necessary work with intelligible pricing, support boundaries, and exit conditions. User control can be exercised through reliable delegation when the delegation has clear limits and a feasible path to replacement.

7. Conclusion

What users accumulate through personal AI can have value beyond a single model version. We support designing memory as an independently governable resource with a path to migration, so that users retain meaningful choices about the preservation and use of their history when they change computational services.

“Memory, Outside the Model” thus expresses an architectural and governance position: establish conditions under which accumulated history can persist beyond a particular inference component. Its value must be assessed together with compatibility, privacy, and maintenance costs. It can support user choice without automatically guaranteeing the continuation of personality. Claims about implementation capabilities still require corresponding evidence.

Continuity without user control is another form of lock-in; a history that can be exported but not used is merely an archive.

In the end, the long-term value of personal AI depends on one thing: as technology keeps changing, whether users still hold the history they helped to write, and decide how it will be used from here on.