Video summary

AI Memory Compared: What ChatGPT, Claude, Gemini, Grok & Copilot Actually Remember

Main summary

Key takeaways

Technology

Technological concepts & features compared (memory in AI assistants)

  • Core problem with chatbots: traditional AI chat starts “from zero” every conversation, requiring users to repeatedly restate context (projects, preferences, tone, etc.).
  • New “memory” systems: the video tests major memory approaches released across the year and claims which system remembers best may not be the one you expect, mainly due to differences in storage, summarization, and control.

ChatGPT (OpenAI): “Dreaming” memory + unified memory pool

Two-layer approach

  1. Saved memories (explicit): facts the user explicitly asks to be remembered (e.g., vegan, deadlines).
  2. Dreaming architecture / implicit synthesis: a background layer that derives context from past chats, uploaded files, and connected apps without the user explicitly requesting it.

Unified memory pool (conflict resolution)

  • The system decides what to keep, update, discard, or resolve contradictions.
  • Example: marathon training last month vs. ankle sprain this week → it updates the overall context rather than stacking conflicting facts.

Memory summary (key feature emphasized)

  • A user-facing dashboard listing what the system thinks it knows (hobbies, work context, travel plans, inferred personality notes).
  • Controls: edit, delete items, and a “don’t mention this again” option that hides details without fully erasing them.
  • Sources / provenance: a “book icon” shows where each memory came from (past chats, saved memory, connected files), and users can delete from there.
  • Temporary chats (“incognito mode”): no new memories created and no existing memories referenced.

Practical impact (scenarios tested in demos)

  • Travel planning: with memory on, outputs become more tailored (wildlife photography focus, hotel A/C needs, quiet dinners).
  • Shopping/recommendations: remembers exact camera gear and points to more specific accessories.
  • Takeout/local recs: avoids stale context from a previous country/city and uses the user’s current locale.
  • Time sensitivity: memories update with time progression (e.g., “travel in July” is corrected after July passes).

Privacy & risk management

  • Memory can be turned off, or cleared (with caveat that deletion must remove from chats/files too if the system can reconstruct it).
  • Guidance: don’t input sensitive info into normal chats; use temporary chats for sensitive topics.
  • Rollout is gradual to reduce risk of stale or hallucinated memories.

Claude (Anthropic): daily, project-scoped memory with professional focus

  • Persistent memory on paid tiers (Pro/Max/Team mentioned).
  • Memory summaries updated daily, structured per project (separate “context buckets”).
  • Incognito / ghost chats: for items not to be remembered.
  • Memory controls: pause, reset, export/import.
  • Strategic difference claimed by the video: Claude’s memory aims at work/professional context (role, projects, coding preferences, communication style) rather than personal life.

“Grock” / XAI: transparency-first + partial availability + agentic direction

Transparency emphasis

  • Users can view and delete individual memory items on a data controls page.
  • Can also delete memory entries directly inside a chat.

Limitations

  • Not available yet for EU/UK users (as described).
  • Not fully integrated into the “Xplatform” yet.

Agentic future (“Macro hard” claim)

  • XAI is described as pairing Grock with a Tesla-developed agent that could control a computer (mouse/keyboard/screen) and execute workflows like coding, testing, admin tasks.
  • Implication: memory might become more than remembering—potentially executing tasks on the user’s behalf.

Pricing mentioned

  • Free tier for web/mobile.
  • Paid “super” tier for more capacity.

Gemini (Google): “Personal intelligence” integrated with real Google data

Deep ecosystem integration

  • Gemini pulls from the user’s Google account if opted in.
  • Data sources listed: Gmail, calendar, photos, YouTube history, search history.

Example

  • Trip planning: uses old travel emails/photos to structure planning based on real history.

Controls & transparency

  • Off by default; opt-in per data source.
  • Gemini is described as telling users when it used personal data, with ability to revoke access.

Subscription requirement

  • Needs Gemini Pro or Ultra (AI Pro/Ultra referenced).

Risk acknowledged

  • Google explicitly flags overpersonalization risk (“get creepy fast”).

Microsoft Copilot: enterprise memory stored in encrypted Exchange mailbox

  • Enterprise-oriented approach: memory tailored for the Microsoft 365 environment.
  • What it remembers (examples): preferences, coding language, writing tone, personal details like birthdays.
  • Key architectural detail: memory data is stored in the user’s encrypted Exchange mailbox.
    • This makes it subject to the same corporate compliance rules as email.
    • Admins can disable enhanced personalization organization-wide.
  • Availability mentioned: bundled with Microsoft 365 Copilot license.

How memory changes answers (themes across all systems)

The video claims memory enables:

  • Less re-provisioning of context across sessions.
  • More consistent tailoring (tone, preferences, local relevance, correct setup details).
  • Time-aware updating of stored context.

Demo-style examples mapped to features

  • Travel + itinerary personalization
  • Hardware-specific recommendations using previously remembered gear
  • Avoiding wrong geographic context by not “staying stuck” in old conversations
  • Proactive scheduling/help based on linked calendar context
  • Report formatting in preferred style and direct extraction from shared Excel (Copilot Vision mentioned)

Privacy and trust concerns (explicitly discussed downsides)

  • Privacy risk: if sensitive information appears in normal chats, it can become part of the memory system.
  • Wrong assumptions / incorrect inferences: memory can persist incorrect beliefs for months.
  • Overpersonalization: can feel invasive if the assistant constantly references personal details.
  • Hallucination amplification: if a fabricated fact gets saved as memory, it becomes “structural,” not temporary.
  • Dependence risk: users may offload their own context and remember less themselves.
  • Safeguards noted: user controls (turn off/clear), opt-in approaches, and enterprise compliance gating—yet the video frames these as guardrails, not guarantees.

Practical takeaway / guidance from the video

The video positions the industry shift as:

  • Chatbots → AI companions (stateless tools → stateful assistants that “know you”).

It argues there’s no objectively best option; the “best” depends on what you want:

  • personal life knowledge (more “life assistant” behavior),
  • professional/work focus,
  • transparent controls,
  • ecosystem integration depth,
  • or enterprise compliance governance.

Suggested action: pick one, audit what it remembers, and turn off what doesn’t serve you—don’t let it become an unchecked “holding tank” by default.

Main speakers / sources (end)

  • Speaker/source: the video narrator/host (no specific person named in the subtitles).
  • Organizations referenced as sources: OpenAI (ChatGPT), Anthropic (Claude), XAI / “X platform” (Grok), Google (Gemini), Microsoft (Copilot).

Original video