Long-term memory in chat is useful—and a new responsibility

MemoryBank sits outside the base model

The mechanism supplements language models with an external memory. Relevant content is selected from earlier interactions, summarized, and provided again in later responses.

This allows the same approach to support different model providers. The memory is not limited to a specific model architecture but is part of the surrounding application.

This separation is technically useful. A base model can be swapped while the profile and conversation continuity are preserved. At the same time, the application must be responsible for the accuracy and protection of the memory.

A more powerful model does not repair a false memory that is fed to it as a presumed fact in the context. Memory quality remains a separate layer.

Memories Are Strengthened or Forgotten

MemoryBank orients its update mechanism to the Ebbinghaus forgetting curve. The passage of time and relative significance influence which content is retained, reinforced, or made less accessible.

Selective forgetting prevents every old message from permanently holding the same rank. For long-term companionship, this makes sense because a complete trajectory becomes unwieldy and costly.

The assumed significance, however, is not neutral. An algorithm can over-weight a casual remark or let a boundary that matters to the person fade away.

Users should therefore be able not only to delete all data, but also to see and correct which persistent settings or memories the system actually uses.

SiliconFriend Demonstrates Long-Term Companionship

The authors deployed MemoryBank in an LLM-based companion called SiliconFriend. The system was additionally fine-tuned on psychological dialogue data.

In qualitative real-world dialogues and quantitative simulated trajectories, it demonstrated the ability to retrieve relevant memories, take the user’s personality into account, and respond more empathetically.

For the simulation, ChatGPT played various user characteristics and generated long-term conversations across many topics. Such data enable controlled variation, but they only partially reflect real human development.

A successful companion demonstrator is not yet evidence of psychological efficacy or safe long-term use. Relationship dynamics and dependency in particular require real, longer-term research.

Memory must separate facts from interpretations

A display name, a chosen tone, or the companion’s name are clear preferences. A statement like “You always feel abandoned in relationships,” by contrast, is a far-reaching interpretation.

When a system stores such interpretations as permanent memories, later responses can repeatedly confirm the assumption. An uncertain impression becomes a seemingly stable personality profile.

The memory should therefore include origin and certainty. Direct user statements, temporary conversation states, and model hypotheses require different rules.

Corrections must not merely supplement older entries but, where applicable, replace them. Otherwise, the refuted assumption resurfaces with every retrieval.

Contradictions between two user statements are not always errors either. People change their minds or describe the same situation differently on different days. The memory should retain temporal context rather than quietly declaring one version the definitive truth.

For sensitive hypotheses, deliberately not storing them can be the best decision. The chat may address them in the current conversation without turning them into a permanent profile feature.

Uncertainty is particularly easily lost during compression

A preprint by Alex Kwon asks not only which statement remains in memory, but whether its epistemic stance survives compression. In the experiments, the actual content remained the same; what varied was whether certainty or tentativeness was noted as an explicitly labeled field or only as an inserted hint.

Across over 60 claims in seven linguistic registers, a separate, clearly labeled status field increased the retention of the epistemic stance by approximately 15 percentage points in two models. A preregistered replication with Claude Haiku showed an increase of 15.6 percentage points. Longer notes alone did not reliably help; individual wording effects also differed between models.

For chat memory, this is an important distinction. The memory "person suspects X" must not become "X is true" during summarization. Source, certainty, temporal reference, and later corrections should therefore remain structurally separate from the claim, rather than merely being mentioned somewhere in a longer note.

The study uses synthetic notes, few models, and measures retention, not the quality of real long-term conversations. As a single preprint, it does not yet establish a reliable memory architecture. However, it provides a testable indication that explicit structure may be more important than merely more text.

A suitable product test would therefore need to ask not only whether information is still present after compression. It would also need to check whether a conjecture, a confirmed statement, a contradicted claim, and an uncertain conclusion retain their distinct statuses.

• Store claim, source, and degree of certainty separately.
• Treat corrections as a change in state, not merely as an additional sentence.
• After each compression stage, check whether uncertainty was preserved.
• Do not infer any claim about conversation quality or safety from retention alone.

Retrieval can reduce hallucinations

Shuster, Poff, and Chen investigated retrieval architectures for knowledge-grounded dialogue. They combined retriever, ranker, and encoder-decoder components to generate responses that are both informative and conversational.

The best models achieved strong results on two knowledge dialogue tasks and, according to human evaluations, substantially reduced knowledge hallucinations.

The principle resembles memory retrieval: rather than having the model recall solely from its parameters, it receives selected external information. Quality then depends heavily on selection and ranking.

Retrieval does not prevent errors if the wrong memory is retrieved or a correct source is misinterpreted. It shifts part of the quality work into storage, search, and context assembly.

Emotional consistency is a different memory task

The Emotional Chatting Machine by Zhou, Huang, and Zhang modeled emotion categories, an internal emotional state, and an external emotional vocabulary.

It could generate responses that were more appropriate in content and emotional expression. This memory primarily controls the form within the dialogue, not long-term personal facts across multiple sessions.

A real companion system may need both levels: momentary conversation dynamics and enduring preferences. They should not be mixed into a single opaque profile.

A momentary desire for quiet must not become a permanent attribution of a personality trait. Conversely, a long-term chosen direct address should not be lost after each session.

Data protection begins with the question of what is not stored

Long-term memory increases utility because people do not have to start from scratch every time. It also increases the amount of sensitive information that remains associated with an account or device.

Data minimization requires a decision about purpose. Which details does the chat really need for continuity, which can remain local, and which should disappear after the session?

Registration and guest mode can have different storage commitments. A guest conversation can remain in the browser; a voluntary account can enable storage across devices. These differences must be understandable before the decision.

Technical encryption is important, but not sufficient. Access, export, deletion, retention period, and use for quality analysis are part of the complete description.

A good memory is controllable continuity

MemoryBank demonstrates a flexible way to build long-term memory around various language models. Its demonstration suggests that relevant recollections can improve companionship and personalization.

For sensitive conversations, however, memory is not a biography that is as complete as possible. It should retain a few useful, verifiable pieces of information and more easily forget what is uncertain.

Tests must examine retrieval as much as storage: Is the right memory used at the right moment, does a deletion remain effective, and does a corrected assumption actually disappear?

Another test concerns account switching. Memories must never be mixed between people, browser profiles, or household devices. Even a single foreign personal memory can destroy the entire space of trust.

Long-term memory in chat is useful and a new responsibility. Only transparency, user control, data minimization, and error-aware retrieval turn memory into helpful continuity rather than an invisible psychological archive.