Memory and new controls for ChatGPT

Original Article ↗ Hacker News Discussion ↗

OpenAI’s new “memory” feature for ChatGPT, which lets the model remember user preferences and context across chats, is seen as both a long-awaited usability boost and a potential lock‑in and privacy risk. Commenters split between “transactional” users who want clean, stateless sessions for security and reliability, and “relationship‑driven” or project‑based use where persistent context would save constant re‑explaining. Many also question how the feature is technically implemented, how it interacts with already “lazier” coding behavior in recent GPT‑4 variants, and why core needs like better search, larger context windows, and a more robust UI are not being prioritized instead.

Usage patterns & value of memory

  • Many distinguish between “transactional” use (each chat a fresh query, like search) and “relationship/topical” use (ongoing projects or quasi‑companion).
  • Transactional users often prefer a clean slate for security, predictability, and debuggability; they worry memory will introduce hidden state and odd behavior.
  • Relationship/topical users want persistent context for projects, preferences, and recurring topics (e.g., coding style, devices used, lesson-planning habits).

Desire for scoped / structured memory

  • Several want memory scoped by topic, project, or “folders,” not blended across the whole account.
  • Requests for tree/threaded conversations, context hierarchies (work vs personal), and reusable “starting frames” from past chats.
  • Some imagine “macro” or profile systems: different base prompts/tones for blogging, technical research, etc.

Implementation details & technical guesses

  • Many assume a RAG-style approach: short “memories” stored separately, queried and injected into prompts when relevant.
  • A prompt leak shows a bio tool used to persist user info; memories appear as short text blobs appended to the system context.
  • Debate over whether this is just refined custom instructions vs something more dynamic; consensus: incremental but practically useful, not revolutionary.

Privacy, trust, and lock‑in

  • Persistent memory increases concern about data retention, training use, and verification of OpenAI’s policies.
  • Some view it as a lock‑in/“stickiness” mechanism, raising switching costs vs competitors.
  • Several prefer local LLMs for privacy, control, and avoiding safety filters.

Model behavior & “laziness”

  • Strong thread on GPT‑4 Turbo becoming “lazy” for coding: partial implementations, comments instead of full code, ignoring explicit instructions.
  • Benchmarks and anecdotes suggest this worsened over recent updates; some think it’s cost-driven or prompt-driven, not just safety.
  • Workarounds mentioned: unified diffs, system prompts emphasizing expert‑to‑expert tone, careful prompting, or switching to other models/locals.

UX, interface, and missing basics

  • Complaints that ChatGPT’s web UI is slow, buggy, and lacks critical features like global search over chats.
  • Some would trade new “memory” and personalization features for core improvements: reliability, larger effective context, and better adherence to instructions.