Home
Apr 15, 2026
View All

Sharing AI Context Across Devices, Character.AI's Memory Update, and Copilot in Word

·2 underrepresented voices

The Short Version#

Three distinct signals today: Teresa Torres shares practical patterns for managing AI context persistence across devices (a real friction point most teams ignore), Character.AI ships a meaningful model + memory update that shows how consumer AI products are raising the floor on personalization, and Microsoft quietly makes Copilot in Word feel more like a coworker than a chatbot. The common thread: the tooling is maturing past "AI can help" toward "AI that actually knows you and your work."

Teresa Torres — How to Share AI Context and Skills Across Devices#

Source: https://www.producttalk.org/how-to-share-your-ai-context-and-skills/ Credibility: High (first-person practitioner post from recognized PM thought leader who hosts Claude Code office hours)

What happened: Teresa Torres published a practical post on one of the most-asked questions from her Claude Code office hours: how do you share AI context and skills across devices? The excerpt signals she's been running these office hours for months and has observed a consistent pattern — people are building up AI workflows on one machine and hitting a wall when they switch to another. The post (paywalled audio version also available) addresses this directly with concrete approaches.

Key PM craft patterns:

  • The "context and skills" framing is important: context means what the AI knows about you, your work, and your codebase; skills means the reusable instructions and prompts you've trained it to use well
  • Sharing these across devices requires intentional artifact management — project files, CLAUDE.md-style instruction documents, or shared config that travels with you
  • This is a solved problem at the individual level but almost no teams have standardized it — which means most AI productivity gains are locked to one person's one machine
  • The fact that this is the #1 question in her office hours tells you something about where practitioners actually are: not "should I use AI" but "how do I make this not fragile"

Why it matters for PMs: This is the friction point that separates individual AI productivity from team AI productivity. If your team is getting value from AI-assisted workflows, ask yourself: what happens when someone changes laptops, adds a new team member, or picks up a task someone else started? If the answer is "they start over," you have a fragility problem that's costing you the compounding benefits of AI assistance. Teresa's framing of "context and skills" is a useful mental model for PM teams thinking about how to build durable AI workflows — not just for engineers, but for anyone using Claude Code, Cursor, or similar tools.

Critical questions:

  • How much of your team's AI productivity is locked inside individual setups that would disappear if that person left?
  • Is there a meaningful difference between "sharing context" and "creating a shared AI onboarding doc"? If so, what is it?
  • At what team size does informal context-sharing break down and require a structured solution?
  • How do you version-control AI context the same way you'd version-control code?

Action you could take today: Audit your own AI setup — if you switched to a new machine tomorrow, what would you lose? Write it down as a "context and skills" inventory. That list is the starting point for a team-level standardization conversation.

Character.AI — New Model, Memory, and Lorebook#

Source: https://blog.character.ai/pipsqueak2-and-more/ Credibility: High (first-party product announcement from tracked company)

What happened: Character.AI shipped their April product update, which includes three meaningful changes: a new model (Pipsqueak 2), an upgraded memory system, and a new feature called Lorebook. This is a notable cluster — they're simultaneously improving the underlying model, making it more persistent (memory), and giving users more control over how characters behave and what they know (Lorebook). The combination signals a maturation from "chat with a character" to "build and maintain a character relationship with continuity."

Key capabilities:

  • Pipsqueak 2: Updated character model — based on the naming convention, this is their smaller/faster model optimized for character interaction, not their heaviest model
  • Memory: The upgrade here is about persistence — the AI remembers things from past conversations, which is the core of what makes consumer AI feel like a relationship vs. a search box
  • Lorebook: A worldbuilding/character-context tool that lets creators (and presumably users) define the rules, lore, and context that govern a character's behavior — think of it as a structured system prompt layer that non-technical users can manage

Why it matters for PMs: Character.AI has always been the clearest example of consumer AI retention driven by relationship, not utility. Memory is the mechanism that makes that work — without it, every session is a cold start and users lose the feeling of continuity that keeps them coming back. The Lorebook feature is interesting for a different reason: it's essentially user-facing prompt engineering made legible. If you're building any kind of AI product where users can customize behavior, watch how Character.AI's Lorebook UI evolves — it's a live experiment in how non-technical users want to control AI context. This directly touches the open question about how much autonomy users want to give AI, and how much control they want to retain.

Critical questions:

  • Does memory actually drive retention, or is it a feature that sounds good but gets ignored by most users?
  • Lorebook is powerful for creators — but does it add friction for casual users who just want to chat? How do you surface it to the right people without overwhelming others?
  • How does Character.AI handle memory conflicts (things the user said that contradict each other across sessions)?
  • Is Pipsqueak 2 better enough to matter for the use cases Character.AI users actually care about?

Action you could take today: Spend 15 minutes with Character.AI's Lorebook — specifically look at the UI for how they've made structured context-setting accessible to non-technical users. If you're building any product where users configure AI behavior, this is a live reference design worth studying.

Microsoft Copilot in Word — Tracks Changes, Leaves Comments, Works Like a Coworker#

Source: https://www.linkedin.com/posts/satyanadella_new-in-word-copilot-now-tracks-changes-activity-7449881077600641024-92HA Credibility: High (direct announcement from Satya Nadella, CEO)

What happened: Microsoft shipped a meaningful Copilot update for Word: it now tracks changes and leaves comments the same way a human collaborator would, grounded in enterprise context via something called "Work IQ." This is a positioning shift as much as a feature update. The framing — "working more like a coworker right inside your document" — is deliberate. They're moving away from "AI generates text" toward "AI participates in your document the way a colleague does."

Key capabilities:

  • Track changes: Copilot's edits show up as tracked changes, meaning you can see exactly what it changed and accept/reject them individually — the same workflow you'd use with a human editor
  • Comments: Copilot can leave inline comments, not just make edits — this is huge for the review use case
  • Work IQ grounding: Copilot's suggestions are grounded in enterprise context (presumably calendar, org chart, prior docs, email) — not just the current document
  • This is available in Word, implying the underlying architecture change is in how Copilot participates in document-level workflows

Why it matters for PMs: Track changes is the interface metaphor that makes AI edits feel safe. When AI generates a paragraph and replaces yours, users lose control and trust erodes. When AI's edits show up as track changes, users retain the final call on every word. This is the right UX pattern for any AI writing feature where the human needs to stay accountable for the output — legal docs, comms, PRDs, anything with a real author. If you're designing an AI writing or editing feature, tracked-change-style UI should be on your list of patterns to evaluate. It's also worth noting that "Work IQ" as a grounding layer is Microsoft's answer to the context persistence problem Teresa Torres is writing about — but at the enterprise infrastructure level rather than the individual practitioner level.

Critical questions:

  • Does grounding in Work IQ actually improve suggestions, or is it a marketing label on existing features?
  • How do users feel when Copilot's tracked changes pile up alongside human collaborators' changes? Is the visual noise manageable?
  • Does the track changes pattern work for generation (new content) or only for editing (modifying existing content)?
  • What happens when Work IQ surfaces confidential context that shouldn't appear in a shared doc?

Action you could take today: If your team uses Word for any collaborative docs (specs, comms drafts, strategy docs), pilot this with one real document this week. Specifically test whether track-changes mode changes how people feel about accepting Copilot's edits versus the previous "just apply it" model.

Quick Hits#

  • Mistral AI: Shipped "Connectors" in Mistral Studio — built-in and custom MCPs with direct tool calling and human-in-the-loop approval controls. A concrete implementation of MCP with approval gates built in, worth watching for anyone thinking about enterprise AI tool access patterns (2026-04-15): https://mistral.ai/news/connectors

  • Dan Shipper: "Software Engineering in 2026 Needs Two Roles: A Pirate and an Architect" — the pirate codes fast to find what's valuable, the architect turns it into something maintainable. Clean mental model for how AI-assisted dev teams are splitting responsibilities (2026-04-13): https://x.com/danshipper/status/[post from Apr 13]

  • Amjad Masad / Replit: Replit shipped configurable app hosting regions, specifically flagged as useful for compliance and privacy laws. Platform expansion into compliance use cases signals Replit taking enterprise requirements more seriously (2026-04-13): https://x.com/amasad/status/2043785145606656223

  • Julie Zhuo: Published "How to Spot a World-Class Designer" — no excerpt, but Zhuo's writing on design craft consistently surfaces frameworks relevant to product decision-making. Worth reading if you're hiring or evaluating design quality on AI products (2026-04-14): https://joulee.medium.com/how-to-spot-a-world-class-designer-1c4c61adb7ce

  • Notion 3.4 part 2: Shipped April 14 — no excerpt in collected data, but Notion has been on a consistent release cadence with AI features. Worth checking the release notes for anything in the AI workflows category (2026-04-14): https://www.notion.so/releases/2026-04-14

The Thread#

Context persistence is the unsexy infrastructure problem nobody's solved. Teresa Torres is writing about it from the individual practitioner angle (how do you keep your AI skills when you switch devices), Microsoft is building it at the enterprise level (Work IQ grounds Copilot in your org's context), and Character.AI is shipping it as a consumer retention feature (memory across sessions). The underlying problem is the same: AI is only as useful as what it knows about you, and right now that knowledge is fragile, siloed, and session-bound. Teams that figure out how to build durable, portable AI context — whether through shared config files, enterprise grounding infrastructure, or explicit memory systems — will have a compounding advantage over teams that rebuild from scratch every time.

Sit With This#

Microsoft's Copilot in Word now shows its edits as tracked changes — the same interface pattern you'd see from a human editor. The stated goal is that it should feel "more like a coworker."

For your AI product: If you have an AI feature that modifies, generates, or suggests content, does your UI make it easy for users to see exactly what the AI changed and accept or reject it piece by piece? Or does it present AI output as a fait accompli? What would it cost to add a tracked-change-style interaction layer — and what would it unlock in terms of user trust and adoption?