A template determines where information goes. A writing preference determines how that information is expressed. Neither should decide what happened in the consultation. Keeping those jobs separate is the starting point for configuring Tandem's Templates and Memories without turning a convenient default into an unsupported clinical statement.
Tandem's 20 April 2026 explanation distinguishes Template Builder from personal and clinic Memories. Templates set the structure; Memories supply reusable wording or formatting instructions. The company describes Memories as stored text read during generation, not changes to model weights. That explains the design without proving that every resulting note follows every instruction correctly.
Begin with the record you need to read
Before changing a setting, write down what makes a note difficult to review. Perhaps the assessment is buried in a long history. Perhaps the plan combines completed actions with possible next steps. Perhaps abbreviations are understandable within one team but confusing to others.
Each problem points to a different intervention. Missing headings suggest a structural change. Excessive repetition suggests a writing instruction. A factual error requires correction and investigation, not merely a more attractive layout.
Take a fictional community clinic where follow-up responsibilities are often scattered through the note. A useful structure might separate the clinical assessment, agreed actions and outstanding arrangements. Those headings make responsibility easier to inspect. They do not mean every encounter necessarily contains an action in every category.
Build the structure before refining the voice
Start from the smallest template that captures the service's actual information needs. Include sections that have a clear purpose and avoid adding fields solely because another speciality uses them.
A heading called "Examination" can be useful, but it must not encourage a normal examination to appear when none occurred. Similarly, a field for follow-up should not become an automatic claim that an appointment has been booked.
For a first test, use an original fictional encounter with deliberately incomplete information. Check how the configuration handles an absent examination, an unresolved question and a plan awaiting agreement. The right result is not the fullest-looking note. It is a note that faithfully represents the available information.
Write preferences that alter expression, not reality
Good candidate instructions describe presentation: use UK spelling, keep the assessment concise, separate historical information from today's findings, or preserve uncertainty when it was expressed. These examples are editorial suggestions, not guaranteed commands with identical effects in every configuration.
Poor instructions supply facts regardless of the encounter. "Always record a normal examination" is not a style preference. "Always state that the patient agreed" could turn an unfinished discussion into apparent consent. Even a well-intentioned request to include every safety-netting point can be misleading if those points were never discussed.
A useful test is to ask whether an instruction would remain true in a consultation where the relevant information was absent. If not, rewrite it so the output depends on what was actually established.
Personal and shared settings need different ownership
As described in Tandem's April announcement, clinicians can set personal Memories and accept suggested ones, while administrators can set shared preferences. The personalisation page checked on 6 September 2026 says Memories can be viewed together, edited and removed.
That does not eliminate the need for local ownership. A personal preference for short paragraphs may be harmless; a clinic-wide convention about documenting follow-up affects everyone. Give shared instructions a named owner and a clear reason for existing.
Where a personal instruction appears to conflict with a clinic default, inspect the applicable settings and confirm the precedence with support. The public material reviewed here does not establish a universal conflict-resolution rule that can safely be assumed for every configuration.
Suggested Memories deserve a deliberate decision
A repeated edit may indicate a useful preference, but it can also reflect an unusual run of consultations. Before accepting a suggestion, ask whether it describes your general writing style or a temporary need.
Suppose a clinician repeatedly shortens a section during brief administrative encounters. Applying the same preference to complex assessments might remove detail that matters. The fact that an instruction saves time in one context does not establish that it belongs in every template.
Read the suggestion as an instruction that may influence future drafts. Check for unwanted specificity and avoid saving patient details as preferences. Tandem describes screening suggestions for patient information, but a product safeguard should not replace careful review of what is being accepted.
A small regression check before wider use
Create three fictional encounters for testing: one uncomplicated review, one with several concerns, and one where the clinician deliberately leaves a question unresolved. Keep the reference accounts fixed while changing the configuration.
Compare drafts before and after the change. Look for the intended improvement, such as a clearer plan, but also check what disappeared or became more certain. A shorter assessment is not better if it loses the reason an investigation was deferred.
Record the template version and Memories used. This gives the team something concrete to restore if the revision causes problems. No product test was performed for this article; the exercise is a suggested local check, not a report of measured Tandem behaviour.
Remove instructions that no longer earn their place
Over time, a collection of individually sensible preferences can become difficult to understand. Two instructions may repeat the same request, or one may have been added to correct a problem that no longer exists.
Review the list periodically. Keep instructions that address a current need, combine duplicates and remove obsolete ones. After a significant change, repeat the fictional cases rather than assuming that a simpler settings page automatically means a better note.
When a draft is wrong, establish whether the issue arose from missing input, transcription, generation or configuration. Adding another instruction is not always the appropriate response.
Keep learning records separate from consultation templates
This guide is published by iatroX and includes its professional-learning tools as a complementary option. It does not suggest that iatroX controls Tandem settings or replaces a clinical record system.
According to iatroX's September 2026 product information, its CPD tools let clinicians review and personalise a draft learning record and export completed evidence. That is a different document with a different purpose. A reflection on improving documentation should describe the learning without reproducing identifiable encounter material.
For an individual clinician, begin with a small number of clear preferences. For a clinic administrator, prioritise shared structure, ownership and testing. For both, the governing rule is the same: personalise the wording, not the facts.
Frequently asked questions
Are Tandem Memories the same as training the model on my notes?
Tandem's April 2026 description says Memories are stored text instructions used during generation, not changes to model weights. This explains the feature's stated design rather than every aspect of the service's data processing.
Can I change or delete a Memory?
The personalisation page checked on 6 September 2026 describes viewing, editing and removing Memories. Shared settings may require the appropriate administrative role.
What should I test after changing a template?
Check a straightforward encounter, a complex one and a case with missing information. Confirm that the new format improves readability without adding facts or erasing uncertainty.