Memory Summary should not replace Saved Memories.

This change has been really disruptive for me, and I think it’s the same underlying issue the OP is describing. I wasn’t using saved memories as casual personalization; I was using them as a lightweight configuration layer. For my workflow, saved memories let me store stable operating preferences and workflow rules: how I want context transferred between threads, how I want drafts patched rather than rewritten, how I want the assistant to handle uncertainty, when to ask clarifying questions, and what kinds of tone or framing are counterproductive for me. This regression is especially problematic because Custom Instructions are still severely space-limited. If users had a large, explicit, editable instruction layer, memory could safely become more summary-like because exact workflow rules could live elsewhere. But for complex workflows, Custom Instructions fill up quickly. Mine already have to cover tone, reasoning style, interaction preferences, safety boundaries, project behavior, and editing preferences. Saved memories filled the gap by acting as persistent configuration items. Replacing them with an implicit summary removes that capability without providing an adequate alternative. The old system worked because saved memories were relatively discrete, inspectable, and user-directed. If a memory was wrong, stale, or no longer useful, I could identify the specific item and fix or delete it. The system was imperfect, but I could usually tell what persistent instruction or fact was being carried forward. The new memory summary changes that. A synthesized summary may be useful for casual personalization, but it isn’t a replacement for explicit saved memories. It collapses many separate control points into a lossy, assistant-mediated interpretation. I can no longer tell as clearly what the model actually knows, what it’s inferring, what it’s prioritizing, or why a response is being shaped a certain way. This matters for people who use ChatGPT as a structured workflow tool rather than just a conversational assistant. My setup depends on keeping separate layers: explicit instructions, project-specific context, personal preferences, transient conversation history, and the model’s own inferences. When those are merged into an implicit summary, the system becomes harder to audit, understand, and correct. It’s also made my workflow noticeably worse. I’m seeing more drift toward unwanted personalization behaviors, such as affirmation-heavy, overly agreeable, or facetious responses. Previously I could manage these preferences through saved memories and Custom Instructions. Now, because the system is more synthesized and less directly configurable, I have less confidence that those constraints are being preserved accurately. I don’t object to summaries, but I do have a problem with them becoming the primary control surface. What I’d like to see is explicit saved memories remain first-class editable objects, separate controls for saved memories and chat-history-derived summaries, and a way to pin exact memories so they aren’t rewritten, merged, or abstracted away. I’d also like import/export or version history for memory, along with either much larger Custom Instructions or a separate long-form configuration layer. The current direction may be better for users who want ChatGPT to “just know them” without managing details. But for users who rely on memory as a configuration system, replacing explicit memories with a synthesized summary removes much of the control that made the feature useful. The new paradigm shifts agency away from the user. Saved memories let me configure the assistant directly. The memory summary asks me to rely on the assistant’s interpretation of how it should configure itself. For my use case, that’s a major regression.