Definition

Product-level mechanisms where a chat assistant retains or reconstructs information about the user across separate conversations—distinct from a single long context window or a user-managed document corpus.

Source-Backed Model (Gemini)

According to the author’s conversation with Gemini (source):

  • A cloud-synced User Summary aggregates distilled notes from prior chats (“memory fragments”).
  • New sessions receive that summary as part of context so the model can personalize without shipping full raw history across threads.

Contrast With Other Patterns

  • AI Knowledge Bases: Often user-uploaded documents the model retrieves at query time; oriented around your files, not automatic profiling.
  • LLM Knowledge Bases: A maintained markdown wiki the user (and agent) curates explicitly; transparent and portable compared to opaque vendor summaries.

Token and Quality Tradeoffs

From developer practice notes (原来我一直用错了 Cowork):

  • Vendor Memory (Cowork, Codex) updates and reads incur token on every task—not just a one-time UX win.
  • For one-off tasks, disabling Memory can reduce cost; author reports higher hallucination rates after enabling Memory in some tools.

Practical Takeaway

Cross-session summaries can make assistants feel “aware” of stable preferences and projects. They are useful for UX but are not a substitute for explicit, auditable knowledge bases or wikis when you need provenance and control. Weigh Memory against token budget and error rate for the task at hand.

Current Evidence In Repo

修订：Cursor Agent，2026-06-23。