Digital memory can contribute to continuity of care, but remembering a conversation is not the same as maintaining a reliable clinical record or taking responsibility for a continuing plan. An AI service would need to keep information current, coordinate decisions and make clear who acts when circumstances change. Familiarity alone does not establish continuity.
The attractive possibility is easy to understand. A patient could avoid repeatedly explaining a previous treatment response, a communication preference or an unresolved concern. The harder question is whether the remembered information remains accurate and useful when the patient encounters another clinician or service.
In its Nolla investment announcement of 18 September 2025, General Catalyst described an ambition for continuous, personalised care spanning diagnosis, treatment and follow-up. That is an investor's account of the intended model, not evidence that digital memory has already delivered better continuity or outcomes.
Three meanings of continuity
The 2003 multidisciplinary review by Haggerty and colleagues distinguishes informational, management and relational continuity. These categories offer a useful vocabulary for assessing a proposed AI care service without equating all forms of continuity with repeated contact through the same application.
Informational continuity concerns the availability and use of relevant information. Management continuity concerns a coherent approach to the patient's changing needs. Relational continuity concerns an ongoing therapeutic relationship. Software could contribute to these in different ways, but success in one should not be presented as proof of the others.
A remembered preference might improve communication. A current record might help another clinician understand previous decisions. Neither establishes that someone is coordinating unresolved problems or that the patient has an ongoing relationship with an accountable professional.
The distinction is especially useful when a service describes its assistant as knowing the patient. Knowing facts, retaining a narrative and maintaining a care relationship are different claims requiring different evidence.
A fictional case where accurate memory still needs updating
Consider a fictional patient, Daniel, who previously discussed intermittent headaches with a digital care service. At that time, he described a particular pattern and agreed a plan. Months later, he returns with a concern that he initially describes using the same words.
A system that retrieves the old discussion could save time. But it should not assume that the earlier pattern, interpretation or plan still applies. The new conversation needs to establish what has changed and which historical details remain relevant.
The previous record may be accurate as a record of that earlier encounter. The risk arises when historical truth becomes an unexamined current assumption. This can happen without the system inventing any information.
The appropriate design question is not whether the assistant remembers Daniel perfectly. It is whether it uses the memory to ask better questions rather than to skip questions whose answers may have changed. This is a proposed evaluation scenario, not a report about Nolla's memory architecture.
Some information should expire, remain uncertain or be rechecked
A useful memory system should distinguish stable preferences from time-sensitive clinical information. A preferred communication format might reasonably persist until changed. A medication list or unresolved symptom description requires a different approach.
Dates and sources are important because they explain what a remembered statement actually means. Patient-reported current treatment, an old prescription record and a clinician's interpretation should not collapse into a single undifferentiated fact.
Unconfirmed diagnoses also need to remain visibly unconfirmed. If an early discussion considered a possibility, later summaries should not silently promote it into the patient's established history. Similarly, a previous decision not to proceed should retain its rationale rather than appear as an unexplained omission.
A proposed design might make review status visible and distinguish current, historical, disputed and superseded information. These are suggestions for evaluation, not verified Nolla features. The test would be whether users and clinicians can understand and correct the record without having to inspect an entire conversation archive.
Compare memory with coordination
The following comparison is hypothetical and does not rank named products.
| Proposed service | What could feel reassuring | What still needs testing |
|---|---|---|
| An assistant that remembers every conversation | The patient does not need to repeat their story. | Whether old interpretations are rechecked and another service can use the information. |
| A service with concise records and named plan ownership | The current plan and responsible team are clear. | Whether the record captures the patient's priorities and changes adequately. |
| A combined model | Relevant memory supports a coherent continuing plan. | Whether information remains accurate across clinicians and organisations. |
A conversationally impressive system could still leave the patient to coordinate care alone. A less personable system could provide dependable continuity through clear records and responsive professional support.
This article is published by iatroX and includes its educational role when considering these different tasks. A learning platform is not a substitute for a care organisation maintaining a patient's longitudinal plan.
The difficult test happens across organisations
Suppose Daniel later attends another service. The useful question is whether the receiving clinician can access an understandable account of the current concern, previous decisions and remaining uncertainty, with appropriate authority to use that information.
A long transcript may contain relevant facts without making them easy to locate. A very short summary may omit the reason the plan changed. The transfer needs enough context for the next decision, not the largest possible data export or the most polished narrative.
The GMC's Good medical practice guidance on continuity and records, checked on 5 October 2026, addresses sharing relevant information and making clinical responsibility clear. It provides a UK professional anchor for the discussion, not confirmation that any named digital service meets those requirements.
A proposed continuity assessment should therefore follow the patient through a transfer, not stop at the boundary of the original application. It should examine whether the receiving professional can identify the current plan and whether the patient knows whom to contact afterwards.
Who maintains the overall plan?
A familiar interface can make several separate encounters feel like one relationship. That experience may be valuable, but it should not obscure who is responsible for unresolved concerns.
A service map should identify who reviews new information, who can change the plan and what happens when a clinician is unavailable. It should also make clear when responsibility has transferred and when the patient is being advised to seek separate care rather than being referred within an existing arrangement.
For Daniel, a useful outcome might be a revised plan with a known point of contact. Another appropriate outcome might be transfer to a service able to assess the changed concern. In neither case is a friendly reminder that the assistant remembers him enough on its own.
This suggests a balanced verdict by need. For reducing repeated history-taking, accurate information retrieval could be useful. For interacting problems or changes in management, plan coordination matters more. For a patient seeking a continuing professional relationship, persistent software memory should not be assumed to provide an equivalent experience.
Use longitudinal information to challenge the first interpretation
An educational exercise could show a learner the earlier record, ask what remains relevant and then introduce the new history. The learner must identify which assumptions require rechecking rather than simply repeating the original conclusion.
As described on 5 October 2026, iatroX's reasoning-focused Tutor offers targeted discussion of clinical understanding. A learning session can explore why changed information matters without claiming that iatroX maintains the patient's care record or coordinates treatment between organisations.
The proposed goal is better use of memory, not more memory for its own sake. Continuity becomes meaningful when relevant information supports a current plan and the patient knows that someone can act on it.
The companion iatroX article on error propagation examines what happens when historical assumptions reach new actions. The article on care after the prescription considers how responsibility should continue when the patient moves between services.
Frequently asked questions
Is an AI conversation history the same as continuity of care?
No. It may support informational continuity, but coherent management and an accountable therapeutic relationship require more than retained text.
Should an AI care system remember every historical statement as a current fact?
No. It should distinguish historical, uncertain, disputed and superseded information, and recheck details that could change the present decision.
How should patients judge a service that says it remembers them?
Ask whether the current plan is understandable, whether information can be corrected and who is responsible when circumstances change. Familiar conversation is only one part of the assessment.
Explore how changing information affects clinical reasoning →