Gemini CLI: Memory Is Being Treated As A Reviewable Change

What Changed

Gemini CLI's most important signal is not simply that it has memory. The stronger signal is that memory is being shaped as a reviewable patch. The Auto Memory inbox commit touches docs, settings schema, memory commands, inbox UI, local executor behavior, skill extraction, and evals. That is a full product surface around memory acceptance rather than an invisible note appended to a hidden profile.

The surrounding commits reinforce the same direction: workspace trust in MCP UX, private memory patch allowlists, shell command safety evals, approval-mode awareness for subagents, policy-engine references, sandbox container naming, and structured non-interactive output.

Operator Consequence

Agent memory becomes safer when it is inspectable. A developer should be able to see what the agent wants to remember, why it matters, and whether the proposed memory should enter the project or remain local to the tool.

Bitter Consequence

Bitter should treat Gemini's memory inbox as a useful upstream evidence source. It should record the proposed patch, the approval mode, the workspace trust state, and whether anything was accepted into Bitter's own durable project memory.

This is also a good evaluation target: can Bitter distinguish reviewable memory from ambient memory?