How to approach UX for agentic systems

Agentic UX starts with memory design

Most teams begin an agent interface with chat. But the quality of an agent is often decided one layer earlier: what enters its working context, what persists, and what gets discarded.

The popular shorthand is that a larger context window creates a better memory. The research is less reassuring. In “Lost in the Middle,” model performance was strongest when relevant information appeared at the beginning or end of the context and dropped when the same information sat in the middle. In one condition, GPT-3.5 Turbo scored below its 56.1% closed-book baseline when the relevant document was buried among 20–30 documents.¹

Capacity is not the same as usable attention. A system can technically hold a fact and still fail to use it at the right moment.

More context can create less clarity

Another study added irrelevant details to grade-school math problems. Across the problems that models could originally solve, no more than 18% were solved consistently across every type of irrelevant context.² The extra information did not add knowledge. It added competition for attention.

For product teams, this changes the design question. Do not ask, “How much can the agent remember?” Ask, “What is the minimum context that reliably improves this decision?”

Stored memory ≠ active context
Active context ≠ reliable recall
Reliable recall ≠ good judgment

Four kinds of memory

A useful agent separates memory by purpose instead of pouring everything into one transcript.

Working memory

The current goal, constraints, recent tool results, and unresolved decisions. It should be small, inspectable, and allowed to expire with the task.

Episodic memory

What happened in earlier sessions: decisions, outcomes, corrections, and exceptions. Store the event and its consequence, not an undifferentiated conversation log.

Semantic memory

Stable facts about the user, organization, and domain. Each item needs provenance, recency, and a way for the user to correct it.

Procedural memory

How work should be done: policies, workflows, permissions, and tool instructions. This belongs in governed rules, not in a model’s improvised recollection.

Compression needs a contract

“Summarize the conversation” is not a memory strategy. Compression should be structured around the decisions the agent will make later. A practical memory record can preserve:

• Goal: what outcome the user is pursuing.
• Entities: people, projects, files, and systems involved.
• Constraints: deadlines, permissions, exclusions, and risk limits.
• Decisions: what was chosen, by whom, and why.
• Evidence: the source and timestamp behind each durable fact.
• Open loops: unresolved questions and promised follow-ups.

The raw material can remain in storage for audit or retrieval. The compressed record is what earns a place in active context.

Retrieval is an interface decision

Memory becomes UX when the system decides what to surface. If every layer sees everything, the agent becomes noisy. If each step sees only the evidence relevant to its decision, the agent can become faster and more legible.

A later evaluation across eight long-context models reproduced the middle-position degradation. Randomly permuting documents and selecting a consensus answer improved accuracy by as much as 24% on the tested single-document QA tasks.³

The lesson is not that every product should run the same technique. It is that retrieval order, filtering, and verification are product behavior—not invisible implementation details.

Design the memory controls

Users need more than a generic “memory on” toggle. They need controls that match the system’s actual behavior:

• Before saving: show what will persist and why it may help.
• While using: reveal which memories influenced an answer or action.
• After correcting: distinguish “fix this response” from “update this permanently.”
• When stale: ask for reconfirmation instead of silently reusing an old fact.
• On deletion: make the scope explicit and confirm that derived summaries are removed too.

These controls turn memory from a vague promise into an inspectable product surface.

Measure usefulness, not volume

Token count and number of saved memories are capacity metrics. They do not tell you whether memory helps. A stronger evaluation tracks:

• Retrieval precision: how often selected memories are relevant to the current task.
• Retrieval recall: how often essential context is available when needed.
• Correction persistence: whether a user’s fix survives future sessions.
• Stale-memory rate: how often outdated context changes a recommendation.
• Decision lift: whether memory improves task success against a no-memory baseline.

Anthropic’s long-context experiments found that extracting relevant quotes into a scratchpad improved recall in every head-to-head comparison they reported, though with some latency cost.⁴ That is a measurable design tradeoff: more visible evidence and better accuracy in exchange for time.

The interface is structured delegation

Good agentic UX does not expose the full complexity of retrieval, ranking, and compression. It exposes the parts a person needs to judge the result: the goal the system inferred, the evidence it used, the durable facts it changed, and the action it proposes.

If the experience feels simple, it should be because the system filtered deliberately—not because it hid uncertainty. The future of agentic UX is not an infinite conversation. It is structured delegation supported by selective, correctable memory.

Sources

1. Liu et al., “Lost in the Middle” (2023) — positional effects in long-context use.
2. Shi et al., “Large Language Models Can Be Easily Distracted by Irrelevant Context” (2023).
3. Ravaut et al., “Evaluating Language Model Context Windows” (2024) — eight-model evaluation and medoid voting.
4. Anthropic, “Prompting Claude’s long context window” — quantitative recall experiments.