Gemini Shows Users a Memory Summary, Omni 1.2 Flash Cuts Video to $0.07 a Second, and Bedrock Customers Face a Usage Policy Deadline

Three developments this week touch what an assistant remembers, what it costs to generate media inside a chat, and which terms your deployment silently accepts. Most details come from secondary coverage and aggregators, so treat specifics as the reporters' own and check primary docs before relying on them.

1. Gemini Makes Memory Something Users Can Read

Reports dated November 9 say Google is rolling out a Memory summary in Gemini Apps, a view of what the assistant has learned about the user, with options to edit or delete it. It reportedly appears in the mobile and web apps, and it requires Keep Activity to be on, so users who turned activity saving off will not see it. The rollout date traces back to a single 9to5Google report, and we could not find Google's own announcement, so treat the details as provisional.

The pattern matters more than the product. Memory that users cannot inspect is memory they cannot trust. If your bot stores facts about people, give them a readable list, per-item deletion, and a clear statement of what switches the feature off. Store memory as discrete, attributable items rather than an opaque blob so an edit actually changes behavior.

interface MemoryItem {
  id: string;
  text: string;        // "Prefers metric units"
  source: 'user_said' | 'inferred';
  createdAt: string;
}

// The summary the user sees is rendered from the same items the prompt uses.
const systemPrompt = base + '\nKnown about user:\n' + items.map(i => '- ' + i.text).join('\n');

2. Video Inside Chat Gets Cheaper

Google launched Gemini Omni 1.2 Flash on November 5. Coverage puts 720p generation at about $0.07 per second, described as a 30% cut, with clips raised from a 10-second cap to 15 seconds and multi-turn conversational editing added. That implies a full 15-second clip costs a little over a dollar. The earlier price of roughly $0.10 per second is derived by outlets, not published by Google. One source says requests to the 1.1 model will route to 1.2 from December 3, but it flags that claim as disputed.

A dollar per clip is still expensive for a free tier. Cheaper generation makes video a plausible tool call, but you still need the same controls as image generation: a per-user quota, a pending state while rendering, and a cost estimate surfaced before the call when the user can choose length. Also pin model versions explicitly so an automatic route to a new version does not change cost or output without you noticing.

3. A Policy Date That Binds Your Deployment

Anthropic announced on October 8 that its updated Usage Policy takes effect November 12. The headline change is a bar on sustained, purposeless cruelty toward Claude, but coverage also lists clarifications on tracking people without consent, autonomous hardware control, and unauthorized reselling or proxying. For Amazon Bedrock customers, an AWS Health notice dated October 14 reportedly says continued use on or after November 12 counts as acceptance.

If you ship on a hosted model, a policy change can alter your product's allowed uses without any code change. Assign someone to read provider policy updates, check whether your use case touches the new language, particularly if you proxy or resell access, and make sure your own terms of service pass the relevant restrictions down to users.

Conclusion

Each story is about visibility. Users want to see what the assistant remembers, builders need to see what a generated clip will cost, and teams need to see which terms they have accepted by continuing to run. Treat memory, spend and policy as first-class surfaces in your product, not footnotes.

Suggested visuals: a mock-up of an editable memory list with per-item delete, and a cost-per-clip table comparing 5, 10 and 15 seconds at $0.07 a second.

— Maya