I looked at what’s currently in the ChatGPT system prompt. It turns out there’s a separate hidden memory that can’t be viewed. There’s no tool to write to it, so it’s likely a separate background process for creating summaries. It’s inserted into the prompt as:

# User Knowledge Memories

Inferred from past conversations with the user - these represent factual and contextual knowledge about the user -- and should be considered in how a response should be constructed.

1. The user is highly invested in DevOps / SRE / Platform Engineering topics: they repeatedly discuss IaC (Terraform, Ansible, Helm, ArgoCD, Vault, Teleport, etc.), DevOps topologies, DORA metrics, SRE runbooks, golden paths, platform‑as‑product, etc.; they are building internal frameworks (e.g., “DDD for infrastructure”, “hypotheses for infra”, “policy‑as‑code”, “immutable artifacts”, “observability loops”, etc.)
<<<...another 9 pieces>>>

Apparently, this is part of the mechanism for linking to previous conversations. Although the same request contains the previous 40 conversations with timestamps, titles, and first messages.