Do Home Robots Need Long-Horizon Memory?

That is why long-horizon robot memory is becoming one of the most important home-robot readiness tests. The question is not whether a robot can recognize an object once. The question is whether it can keep a stable, privacy-safe model of a messy home long enough to finish a real task.

In its April 29, 2026 research post, Config demonstrates memory-dependent tasks with a robot foundation model: the information needed for an action has disappeared from view by the time it must act. Config is not selling a consumer home robot, and its demos are controlled research tasks. But the demos translate cleanly into a buyer question: does this robot remember the world, or is it just reacting to the frame in front of it?

What does long-horizon robot memory mean?

Long-horizon memory is the ability to use information from earlier in a task, not just the last few video frames. In a home, that can mean seconds, minutes, or whole routines.

For a vacuum, memory may only mean room maps, no-go zones, and cleaning history. For a companion robot, it may mean people, preferences, reminders, and daily patterns. For a mobile manipulator or humanoid, memory gets much harder: the robot has to track objects, task steps, tool positions, safety constraints, and human instructions while the scene changes.

Config frames the problem well. Its memory post argues that many physical tasks cannot be solved from the current observation alone. A piano-playing robot has to remember which notes it already played. A shell-game robot has to track an object through occlusion. A desk-tidying robot has to remember the original configuration after someone disturbs it.

Those are not chores, exactly. But they are good proxies for chores. Folding a shirt, loading a washer, setting a table, putting toys away, or bringing an object from another room all require the same underlying skill: hold the task state in memory while the world moves on.

What did Config actually demonstrate?

Config's research post and current model page describe different kinds of evidence: controlled task results, model specifications, and the scale of data collection. They should not be read as a single household-readiness score.

As checked on October 11, 2026, Config's CFG-1 model page lists 2 billion parameters, three-minute native memory, and under 50 ms inference on an RTX 5090. It separately lists 10,000+ in-house training hours and 200,000+ collected hours across 10,000+ tasks. These are company-reported figures with different scopes: collected data is not necessarily the data used to train a particular model or evaluate a particular memory task.

Three-minute context also does not establish persistent household preferences, next-day recall, or owner-accessible deletion controls. Those need separate product evidence. A model remembering an earlier observation within a trial is not a demonstration that a purchased robot will remember your instructions tomorrow.

In the April research post, Config reports 70% success on 8-note piano trials and 40% on 12-note trials. In the shell-game test, it reports 21 successful trials out of 24. In desk rearrangement, where the robot had to restore a disturbed scene to a remembered layout, it reports 14 successes out of 20.

These are results on defined research tasks, not estimates of success on household chores or a safety assessment around people and pets. For the home, memory has to be paired with uncertainty handling: ask before acting, pause on low confidence, and leave a task in a safe state when the remembered plan no longer matches reality.

Which home robots are closest to needing memory?

ui44's database makes the distinction clearer. Some robots already use maps and routine memory. A much smaller set is close to the manipulation problem Config is studying.

The important split is between remembering a person or map and remembering a manipulation task. A companion robot can remember that Mom likes a check-in at 9 a.m. A mobile manipulator has to remember that the blue cup was on the left side of the counter before someone blocked the camera, then recover if the cup is now behind a cereal box.

Both use remembered information. Neither memory claim alone establishes successful physical action.

What should buyers look for in a robot memory claim?

Treat "memory" as a stack, not a single checkbox.

1. Map memory means the robot can keep rooms, zones, docks, no-go areas, and navigation routes stable over time. This is common in robot vacuums and mobile monitoring robots. It is necessary, but it does not prove the robot can remember objects or chores.

2. Object memory means the robot can keep track of things even when they are partly hidden, moved, or temporarily out of view. Config's shell-game demo is a clean research version of this. In a home, the test is whether the robot can remember that the TV remote went under the magazine, not hallucinate a new remote because the camera sees a black rectangle.

3. Task memory means the robot knows what it has already done. Did it pour the water? Did it close the fridge? Did it put the spoon down before reaching for the bowl? This matters for every multi-step chore.

4. Routine memory means the robot learns user preferences, schedules, recurring tasks, and household patterns. This is where companion robots like EBO Max are already making explicit claims. It can be useful, but it also creates privacy and consent questions.

5. Safety memory means the robot remembers constraints: do not enter this room, do not touch these objects, do not wake this person, do not share this camera feed, do not repeat a failed grasp without asking. For home robots, safety memory may be the most important layer.

A credible product page should say which layer it means. "Learns your home" is not the same as "remembers hidden objects." "Personalized" is not the same as "can resume a half-finished chore after interruption." "VLA" is not the same as "safe long-horizon autonomy."

Why memory is a privacy feature, not just an AI feature

Memory can retain information about your home briefly during a task or persistently across sessions. Depending on the product, that could include a map, a voice profile, a routine or task history. A model's short context window does not tell you which records a product keeps or for how long.

So the buyer checklist has to include privacy questions:

• Is memory processed on-device, in the cloud, or both?
• Can you view, edit, export, or delete what the robot remembers?
• Are maps, camera clips, voice logs, and task histories separate controls?
• Does the robot train shared models from your home data by default?
• Can guests, caregivers, or family members opt out of recognition?
• Does the robot stop remembering in private zones?
• What happens to memory if you sell, return, repair, or replace the robot?

This is where a robot like SwitchBot onero H1 is relevant because SwitchBot's CES announcement describes on-device OmniSense VLA processing. On-device processing can reduce exposure, but it is not automatically a privacy guarantee. Buyers still need retention controls, local reset options, clear consent, and a boring privacy policy that says exactly what leaves the home.

The safer default is simple: memory should be local when possible, visible to the owner, scoped by room and user, and easy to wipe. A robot that can remember your home but cannot explain or delete its memory is not ready for serious household trust.

What should a supervised memory demonstration show?

Ask a supplier for a supervised, repeatable demonstration of a supported task, including its interruption and recovery procedure. Keep people outside the robot's operating area and use harmless objects. These are questions for a controlled demonstration, not instructions to improvise tests around a moving robot.

For object memory, ask whether the system can distinguish an earlier observed layout from the current one. With the robot stopped and following the supplier's procedure, a demonstrator could cover or move a harmless object. Before any resumption, the system needs a new observation or another appropriate check; an earlier location is not proof that the object is still there. Request an explanation of what happens if the new observation conflicts with memory.

For task memory, ask the supplier to demonstrate an approved interruption between steps of a supported task. A good robot should say what it already completed, what remains, and what changed. A weak robot will restart, duplicate an action, or continue as if nothing happened.

For routine memory, ask whether a supported preference survives a new session or software update, and how the owner can inspect or change it. Keep this separate from the within-task memory demonstrated by Config.

Finally, ask how deletion works for each kind of record, including cloud copies and backups. A demonstration that a robot stops using a preference can be useful, but does not prove that every stored copy was erased. Look for documented controls and retention terms as well as observable behavior.

Where does this leave today's buyers?

For most people, long-horizon memory is not a reason to buy a humanoid in 2026. It is a reason to be more specific when comparing products.

For a home-focused humanoid such as 1X NEO, ask which chores it can resume after interruption, which require scheduled human assistance, and what information it retains. Home positioning and personalization claims do not answer those questions.

For a development platform such as Hello Robot Stretch 4, separate its mapping and software tools from the policy and integration work needed for a particular assistive task. A platform can support experimentation without arriving as a household butler.

If you are a developer, distinguish the standard Unitree G1 from G1 EDU. Unitree lists the standard G1 at USD 13,500 excluding tax and shipping, while G1 EDU requires a sales quote. Its table assigns secondary development and optional dexterous hands to EDU. Confirm that configuration before budgeting for memory-policy experiments; the base price does not establish the cost of a development setup or a household helper.

If you mainly want presence, reminders, or family monitoring, companion robots such as Enabot EBO Max advertise a different form of memory centered on interactions and routines. Just do not confuse social or patrol memory with the ability to manipulate objects safely.

The bottom line

Home robots need long-horizon memory, but memory is not enough by itself. A robot also needs perception, manipulation, error recovery, privacy controls, and a safe way to ask for help when memory is uncertain.

Config's demos are useful because they show the exact gap that polished home robot videos often hide: the important information is frequently not visible when the robot has to act. That is real life. Someone closes a cabinet. A child moves a toy. A pet blocks the path. A caregiver changes the routine. A useful home robot has to remember enough to continue, and be humble enough to stop when its memory may be wrong.

That is the practical buying rule. Do not ask whether a robot has memory. Ask what kind of memory it has, where that memory lives, how long it lasts, whether you can delete it, and which real chores still work after the scene changes.