Memory, inboxes and the links that earn trust

1. Give an agent a memory boundary

Meta introduced Muse Wonder and Muse Wonder Full on October 12. It describes Active Memory as carrying relevant information across conversations. Wonder is available through its model API, a connection developers use to build software, and Muse Code; broader Muse experiences are still to come. These are vendor claims, not independent proof of reliable recall.

Illustrative scenario: an advisor wants an assistant to remember a client’s reporting preferences between monthly reviews. That could save repeated setup. It could also carry an outdated instruction into the next report. Remembering more creates another record to maintain.

This matters in a Canadian operating context where privacy practices vary: the Privacy Commissioner’s 2025–2026 business survey found 64% of surveyed firms with 1–19 employees reported internal staff privacy policies. That finding describes the survey, not every small business.

This week, use fictional client details to test one preference: add it, change it, then ask the assistant what it remembers. Check whether you can inspect and remove it before using real records. Meta’s launch post says further safety evaluations are forthcoming; it does not establish the controls or Canadian availability for your intended setup.

2. Prepare the handoff before buying the inbox feature

Google’s October 15 AI Inbox announcement describes suggested tasks and summaries across email conversations. Rapid Release rollout began October 12, with Scheduled Release starting October 26. Crucially, the stated availability is US users with US English. Starter and Education editions are listed for later. Canadian teams should not plan around immediate access.

Illustrative scenario: a service coordinator returns after two days away and needs the promises, deadlines and unanswered requests. A suggested task list could shorten the catch-up. The tradeoff is that a tidy summary can conceal a changed deadline or an ambiguous commitment.

The first move this week needs no new subscription. Make three fictional message threads: a straightforward request, a revised date and a cancelled order. Write the correct task, owner and deadline for each. Keep this as your acceptance test for any inbox assistant you already have approval to assess.

For a team working across Canada and the US, check user location and language eligibility separately from the subscription name. The announcement is a rollout notice, not a promise of Canadian access or accurate task extraction. Confirm a suggested commitment against the original thread before acting.

3. Make the page behind the answer useful

Google announced clearer links in AI Mode and AI Overviews on October 15. It says inline links will appear higher more often and describes a test that highlights official links with context. An inline link is a clickable source inside the answer itself. A test is not a settled feature.

Illustrative scenario: someone asks whether a Canadian supplier serves their town and how to request a quote. A link helps only if the destination answers those questions. Clear service areas, current terms and a working contact route give a reader something to verify.

The benefit is a more direct path from an AI answer to the business’s own information. The tradeoff is limited control: Google’s announcement gives no quantified traffic gain for small businesses or specific Canadian rollout timetable for these displays.

This week, choose one common pre-sale question. Open the page that should answer it on a phone. Check that the answer, relevant location and next step are easy to find. Ask a colleague to verify the same facts from the page. Treat better source pages as a customer-service improvement; do not forecast extra leads from this announcement.

The bigger pattern

The bigger pattern is continuity: information now travels between conversations, into task lists and out through source links. Each handoff needs a person who can check what survived. Start with one known answer and one exception, then compare the tool’s output with both. Expand only when the result is useful enough to repeat.

Which handoff in your business most often loses a changed instruction: the client history, the inbox task list or the information on your website?