Claude’s memory architecture is the opposite of ChatGPT’s

Attention, addiction, and social impact

• Several comments liken ChatGPT to social media: optimized for attention, potentially harmful to kids and society, and hard to “turn back.”
• Some see an evolutionary split: advantage either to people who exploit LLMs well or to those who avoid the “attention‑sucking knowledge machine.”

User experiences with memory

• Many users disable ChatGPT or Claude memory to avoid unwanted cross‑pollination between unrelated topics, context rot, or resurfacing of hallucinations.
• Others say ChatGPT’s automatic recall is a huge productivity boost, especially for ongoing projects, and is their main reason to keep using it.
• People report ChatGPT inconsistently remembering explicit preferences (e.g., language‑learning settings) while quietly remembering other details like employer and tech stack.
• Some like Claude’s explicit, on‑demand memory but complain that relying on raw history / vector search misses more abstract or personal references.

How memory is actually implemented

• Several commenters argue the article overstates or misinterprets ChatGPT’s behavior, noting:
  • Two memory layers: explicit user memories injected into the prompt, plus embeddings‑based history retrieved via RAG.
  • Recent chats are not fully in context every turn, and the model doesn’t control which snippets are injected.
• Others point out that asking ChatGPT how its own memory works can yield hallucinated implementation details.
• Anthropic’s original “search over raw history” is praised as transparent and controllable; the newly announced enterprise memory that’s closer to ChatGPT’s raises mixed feelings.

Ads, profiling, and business‑model fears

• A strong theme: ChatGPT’s memory and routing are seen as laying groundwork for detailed user profiling, personalized ads, and affiliate links, even if not yet active.
• Some argue ads are economically inevitable given huge costs and lack of current profitability; others counter that subscriptions and enterprise may suffice.
• There’s deep concern that centralized LLM memories will become the ultimate surveillance/profiling substrate, sold to advertisers, employers, insurers, and governments.

LLM understanding, intelligence, and AGI

• Big sub‑thread debates whether “nobody understands LLMs,” with distinctions between knowing the training algorithm vs explaining emergent behavior.
• Another long debate centers on whether LLMs are “just Markov chains,” lack real concepts/world models, and thus cannot reach AGI, versus views that human cognition may be similarly mechanistic and that current models already show some conceptual understanding.
• Skeptics doubt LLMs alone will yield AGI; others expect further architectural innovations (e.g., non‑linguistic, encoded memory, world‑model components).

Privacy, control, and external memory

• Some see “memory as moat” and warn against a future where a few vendors know users better than they know themselves.
• Power users prefer manual context management, APIs, or external stores (e.g., MCP tools) to keep data local and avoid opaque, provider‑controlled memory.
• A recurring practical worry is “context rot”: models learning from their own mistaken outputs if memory is not carefully designed and curated.