AI-native design · Pattern library

Design memory people can trust

Eight interaction patterns for agent memory that stays visible, understandable and under human control.

Explore the reference
The design model

Memory is a relationship, not a database

Agent memory changes future behaviour, so people need more than a list of stored facts. They need to understand how a memory was formed, how strongly it is believed, when it was last checked, and what control they have now.

This library treats provenance, uncertainty, correction, context and deletion as interface material — not implementation detail.

01
Remembered preference

Show provenance

State what was remembered and where it came from. A memory without a source can feel like surveillance.

Do
Name the originating conversation, action, or setting.
Avoid
Presenting a remembered detail as if it were newly discovered.
02
Inferred preference

Show confidence and allow override

Keep an inference visibly different from something the user explicitly told you.

Do
Use tentative language and put the correction beside the inference.
Avoid
Turning repeated behaviour into an unexplained fact.
03
Stale memory

Flag recency

Old memories should ask for confirmation before they quietly shape a new result.

Do
Show the date and offer confirm, update, or remove actions.
Avoid
Using age alone to imply a memory is still true.
04
User-corrected memory

Preserve correction and original

Make the current truth prominent while retaining the superseded value in the audit record.

Do
Label who corrected it, when, and what value it replaced.
Avoid
Showing the original as an active memory after correction.
05
Sensitive memory

Only reveal it in context

Keep sensitive details out of general memory views. Reveal the minimum only when the current task requires it.

Do
Explain why the detail is relevant before revealing it.
Avoid
Using concealed text as decoration — it still exposes that a secret exists.
06
Confidence level

Pair low, medium, or high with a source

Confidence is only useful when people can understand the evidence behind it.

Do
Use words, not colour alone, and link confidence to its evidence.
Avoid
Showing a precise percentage that suggests false certainty.
07
Deletion

Hard delete from inference, soft delete from audit

Stop deleted information from influencing the system while retaining a minimal, non-recoverable audit event.

Do
Describe both outcomes before the user confirms deletion.
Avoid
Keeping the deleted value in logs, embeddings, or summaries.
08
Audit trail

Make it accessible, not buried

Put "View history" beside the memory it explains and provide one clear route to the complete log.

Do
Record creation, use, correction, and deletion in plain language.
Avoid
Hiding memory history inside a generic privacy settings maze.
Put together

A memory surface should explain itself

This interactive reference keeps state, evidence and controls in the same place. Try changing the context, correcting a memory, deleting an inference and opening its history.

Reference implementation

Memory inspector

Current contextGeneral conversation
3 active

What I remember

You are in control
RememberedUpdated today
Prefers aisle seats

You said this in “Sydney trip planning” · Today, 10:42

InferredMedium confidence
You may prefer morning flights

Based on 3 flights you selected · First observed 4 months ago

Check thisLast confirmed 14 months ago
Home airport is Melbourne (MEL)

You set this in Travel preferences · 4 May 2025

Sensitive memories protected
Nothing sensitive is shown here

Relevant details appear only during a task that needs them. Try “Plan a flight” above.

Before you ship

Memory interface checklist