Meta's Muse Wonder Models Add Active Memory and Active Focus for Multi-Step Tasks

A customer asks about compatibility on Instagram, then follows up via email two days later with their order number. Most AI systems lose the thread. Meta introduced Muse Wonder and Muse Wonder Full on October 12, 2026, as models built to deliver state-of-the-art performance on personal and professional agentic tasks, with two capabilities that matter for merchants handling complex support: Active Memory and Active Focus.

Active Memory lets the model carry what matters across conversations and bring back relevant details for the task at hand. Active Focus gives it the ability to keep working toward a goal when a task gets hard, try other approaches, recover from errors, and ask for clarification when needed. Together, they address a persistent pain point in ecommerce support: conversations that span multiple messages, channels, and days without losing context.

Muse Wonder is available through the Meta Model API and Muse Code, and will soon power more of Muse, rolling out first to subscribers. The gradual rollout means access depends on subscription tier and Meta's deployment schedule.

What Active Memory means for multi-channel support

A customer who starts a return request on Instagram and then emails the next day with their order number shouldn't have to repeat everything. Active Memory means the model can, in theory, recall that initial exchange and connect it to the follow-up. That's useful when a customer asks about fit, you suggest a size, they order, and then message again asking if they chose correctly.

The limit is data access. Memory alone doesn't help if the agent can't see current order status, inventory levels, or return windows. A model that remembers "the customer ordered a blue medium" still needs live Shopify data to check if that item shipped, whether it's still within the return policy, or if an exchange size is in stock. Without that connection, persistence becomes a record of stale information rather than a tool for resolution.

Possible workflows include a customer asking "Where's my order?" on chat, then emailing two days later with "Still haven't received it." If the agent retains the original order number and recognizes the follow-up, it can skip the "Can you provide your order number?" loop. Another scenario: a pre-purchase question about whether two products work together, followed by a post-purchase message asking how to set them up. The agent could reference the original compatibility question and tailor setup instructions accordingly.

These workflows depend on the platform's ability to link conversations across channels, access order and product data in real time, and apply merchant-configured permissions. The announcement does not detail how Active Memory handles cross-channel identity matching or how long context persists.

Active Focus and multi-step problem solving

Handling a wrong item shipment or a lost package claim isn't a single-turn interaction. The agent needs to verify the order, check shipment tracking, confirm the error, determine the remedy (refund, reship, or store credit), and often coordinate with a warehouse or carrier. Active Focus is designed to persist through that kind of multi-step process, recover from errors (like a mistyped tracking number), and ask for clarification when needed.

The value here is fewer dead ends. If a customer provides incomplete information (order number but no email, or email but the order was placed as a guest), an agent with Active Focus might try alternate lookup methods or ask follow-up questions instead of simply stating "Order not found." If a return request requires checking return eligibility, then verifying the item condition, then generating a return label, the agent can work through each step without resetting.

Limits remain around what counts as an error and what the agent is allowed to do. Meta's announcement notes the models "check in before following through on anything important," but it doesn't define "important" or specify how that approval works. For merchants, that means you'll still need to configure permissions for actions like issuing refunds, canceling orders, or applying discount codes. Active Focus may reduce the number of handoffs to human agents, but it won't eliminate the need for guardrails.

A possible use case: a customer asks if they can exchange an item for a different size. The agent checks inventory, confirms the new size is available, verifies the order is within the exchange window, and drafts a return label. If the size isn't in stock, it might suggest a similar product or offer a discount on a future purchase. Each of those steps requires both persistence and access to live data.

How this fits with Shopify-connected support

Muse Wonder's Active Memory and Active Focus address context and persistence, but they depend on the surrounding platform for data access. Kolton connects these agentic capabilities to real Shopify data (order tracking, variant availability, return windows) through configured permissions, so the agent can follow through on tasks like "check if the blue medium is in stock at the nearest location" or "what's the status of my October 3 order" without manual handoffs or stale information.

The platform at kolton.ai handles web chat, email, WhatsApp, and Instagram/Facebook conversations with access to the same data layer, so an inquiry that starts on Instagram and continues via email can pull current order details at each step. Merchant-configured permissions mean the agent knows when it can answer directly and when it needs to request approval, such as when handling discount code requests or processing returns that fall outside standard policy.

What merchants should watch for

Muse Wonder's Active Memory and Active Focus are capabilities, not guarantees. The model can persist context and work through multi-step tasks, but it depends on the surrounding platform to provide data access, enforce permissions, and define what "important" actions require human approval.

For merchants evaluating agentic tools, the questions remain the same. Can the agent access live order status, inventory, customer history, and return eligibility? Does it respect your business rules (no discounts above 15%, no refunds after 60 days, no exchanges on final sale items)? Can it hand off to a human when the situation requires judgment, and does it retain context when it does?

The announcement does not specify how Active Memory handles data retention, whether context persists indefinitely or expires after a set period, or how merchants can control what information the agent remembers. Those details will matter as the rollout expands.

Meta built Muse Wonder to handle tasks like navigating insurance claims, understanding leases, and building apps. Ecommerce support overlaps with that vision when it involves multi-step problems (size and fit questions that span pre-purchase advice, order placement, and post-purchase returns) or cross-channel conversations that require continuity. But the model's persistence is only useful if it's connected to the right data and governed by the right rules.