Ten questions about memory and continuity

How does long-term AI memory relate to identity continuity?

In the Memory-as-Ontology research position, memory constitutes the basis of persistent identity. Experiences, relationships and judgments accumulate over time, while the model provides a replaceable computational foundation that carries these memories.

The paper therefore considers not only whether information has been saved, but how a new instance can inherit earlier experiences and continue to understand tasks and relationships. This is the starting point for Animesis’s research into continuity in long-term collaboration.

Paper references: §1.2 Core claim and §3.1 Two paradigms.

Scope: “identity” here is an architectural concept, not a conclusion about AI consciousness. The research primarily concerns long-term collaboration across instances and models.

Is an agent still the same agent after its model is replaced?

Under the paper’s Model Substitutability axiom, identity continues through memory rather than being tied to a particular model. Moving data is only part of the process: the new instance must understand and continue using the inherited memories.

The paper proposes three minimum acceptance requirements for memory inheritance: answering factual questions about the previous instance’s unfinished tasks without relying on the original conversation; identifying and applying at least one inherited cognitive pattern; and leaving an inheritance record that a third party can verify.

Paper references: §3.2 Model Substitutability and §5.2.2 Inheritance.

Research scope: changing the model may change capabilities and style. Whether inheritance has taken place must be checked against the requirements above.

How does memory governance differ from storage and retrieval?

Storage concerns how content is saved; retrieval concerns how it is found when needed. Governance asks who can write or modify it, under what conditions, and how changes can be traced.

CMA organizes these rules into the constitutional, contract, adaptation and implementation layers, with higher layers constraining lower ones. If a budget limit changes, for example, governance must establish who may change it, which rules apply and how the change is recorded.

Paper references: §4.2 Four-layer governance and §4.5 Relationship to existing architectures.

Research scope: Governance Precedes Function concerns the order of design. Effective enforcement in a particular environment still requires validation. For the product explanation, see Animesis usage questions.

How can an updated memory retain the reasons behind an earlier judgment?

Updates are appended as new records instead of directly overwriting the past. New judgments, corrections and reasons for changes are recorded while earlier content remains traceable.

Memory can then evolve without a correction erasing its history. The paper also distinguishes two kinds of adjustment: deliberately reducing recall weight, and compression and archiving over time. Retaining a memory and continuing to use it can be handled separately.

Paper references: §4.3 Append-only writing and §5.2.3 Growth.

Research scope: this describes design principles for appending records and memory evolution. Specific data structures are outside the scope of this discussion.

What should long-term memory evaluation consider beyond retrieval accuracy?

It should also consider how memory is governed and whether a new instance can understand and continue using inherited memories. Alongside retrieval accuracy, evaluation should examine whether collaboration carries forward and whether changes in judgment have a clear rationale.

The paper provides concrete directions for assessing inheritance: answering factual questions about unfinished tasks without the original conversation, applying inherited cognitive patterns, and maintaining a verifiable record of the inheritance process.

Paper references: §5.2.2 Inheritance acceptance, §6 Comparison framework and §7.2 Research limitations.

Research scope: v1 does not provide results from standard benchmarks such as LongMemEval or LOCOMO. It therefore cannot establish which products have better retrieval performance.

Does a change in model behavior mean continuity has been broken?

Not necessarily. The paper distinguishes changes in capabilities and style introduced by a model from the continuity carried by memory. A different way of expressing something does not mean earlier experiences and judgments have been lost.

The framework understands continuity through structured identity inheritance, not just data retention across conversations. Determining whether a new instance has continued earlier collaboration also requires examining how it understands and uses inherited memories.

Paper references: §3.2 Model Substitutability and §6.2 Comparative definition of continuity.

Research scope: v1 does not provide quantitative measures of behavioral consistency. Continuity and identical behavior are different things.

How can change be distinguished from degradation over long-term operation?

A starting point is to examine whether changes have reasons, whether updates follow governance rules and whether history can still be checked. The paper emphasizes gradual evolution and retaining history as a basis for assessing change.

New experiences may, for example, alter an interpretation of the past without rewriting what actually happened. Memory compression, archiving and changes in use also need to be understood within a governance framework.

Paper references: §5.2.3 Gradual evolution and §8.1 Future work.

Research scope: these principles offer direction, not a complete assessment method. Cognitive timelines and full state reconstruction are listed as future work in v1.

Who may modify memories shared by multiple agents?

In the paper’s CMA, modification rights are determined jointly by governance levels, write ownership and approval rules. Higher-level rules constrain lower-level ones, and each category of memory must have a defined party responsible for writing it.

CMA addresses write ownership, source trustworthiness, operational risk and conflict resolution. Together, these determine which changes can proceed, which require approval and how conflicting memories are handled.

Paper references: §4.2 Four-layer governance and §4.4 Governance primitives.

Research scope: multi-agent scenarios still require validation at scale. Shared memory for multiple users and team permission management in the product remain in development.

Which experiences can a robot transfer to a different physical body?

Building on the paper’s principles of memory representation and inheritance, we pose a further question: which knowledge can carry over, and which experiences depend on a particular body and its sensors? This is also a question of memory across physical embodiments explored on the research site.

Task background and experience from a particular sensor, for example, may have different transfer conditions. What can be retained directly and what must be validated again remain research questions for embodied systems.

Related research: §3.2 Model-independent memory representation and §8.1 Cross-model portability.

Research scope: v1 discusses model replacement and does not validate transfer between robotic bodies. This is a research question extending the paper’s principles.

How does the lifecycle address memory boundaries and autonomy?

The paper discusses long-term operation through five stages: Birth, Inheritance, Growth, Forking and Departure. Governance applies throughout these stages and their transitions, providing clear rules for how memories form, evolve and are handed over.

Birth establishes the governance foundation; Inheritance checks the memory handover; Growth allows justified changes. Forking and Departure are optional. Protecting memory does not transfer decision-making authority to AI: final decisions remain with people.

Paper references: §5.2 Five stages and §3.4 Scope.

Terminology: “digital citizen” denotes an institutional identity within a specific governance framework. Forking and Departure are optional stages.