Google's Elephant and Goldfish: A Reviewer That Forgets on Purpose

There is a specific kind of useless feedback that every writer, designer and engineer has received. You hand over a draft. The reader says it looks good. Two weeks later, after the thing ships and breaks, the same reader says they had wondered about that part but assumed you had a reason.
They assumed you had a reason because they were in the room when you decided. Context is what makes a collaborator useful and what makes a reviewer blind.
Google's elephant-goldfish-model, which climbed GitHub's weekly trending list this week past 18,600 stars, is built around that observation. It is an open-source plugin for AI coding assistants with an unusually literal division of labour. An elephant holds your project context and drafts the specification. Goldfish reviewers examine the draft and the source material it references, without the earlier conversation. You make the architectural trade-offs and approve the spec.
The tagline is "Design with an elephant. Review with a goldfish." It sounds like a joke about attention spans. It is actually a design constraint.
What it installs into
EGM runs inside the assistants people already use, which is why its adoption curve looks different from a standalone tool. There is a published plugin for ChatGPT and Codex desktop. For Antigravity CLI the command is agy plugin install https://github.com/google/elephant-goldfish-model; the Antigravity app and the VS Code extension use /plugin instead.
For Claude Code it is two commands:
claude plugin marketplace add google/elephant-goldfish-model
claude plugin install egm@egm
Codex CLI mirrors that shape with codex plugin marketplace add and codex plugin add egm@egm. You start a new session after installing, which is not incidental. The plugin needs the session boundary to keep the reviewer clean.
The 60-second version
You do not need to write a new specification to try it. Point it at something that already exists:
/egm:egm-review docs/designs/webhooks.md
A fresh reviewer reads the design, its clarity, whether it fits the code, and whether the implementation plan holds up. There is also a drift check, which compares a folder of design documents against the current source and reports where they diverge:
/egm:egm-drift-report docs/designs/
The output of that one is deliberately not a verdict. It tells you where the document and the code disagree. You decide which one is wrong.
For non-software documents, /egm-review-loop lets you choose the reviewers and the standard they review against. The built-in draft-and-review workflow targets software specifications specifically, and the bootstrap and drift reporting only apply to codebases.
Why the separation is the interesting part
Most "AI reviews your work" products are one model with a long memory. It reads your draft, remembers the conversation, remembers your stated goal, and then tells you whether the draft meets the goal you both already agreed on. That is a consistency check, not a review. It cannot surface an assumption that was never stated, because the assumption is in its context window and therefore invisible to it.
EGM's goldfish has no context. It gets the draft and the referenced material, nothing else. If the draft depends on a decision that lives only in your head, the reviewer will ask about it. That is the entire mechanism, and it is a better one than it sounds, because unstated assumptions are the most expensive kind of defect. They are cheap to fix at the document stage and brutal to fix after three services have been built around them.

The elephant half matters too, in a quieter way. Having one agent hold the full project context and produce the draft means the draft inherits everything the project already knows. Having a different agent review it means the draft gets checked against what it actually says. Two agents, two jobs, one of them deliberately ignorant.
Where this pattern goes next
Strip the software framing and the pattern is generic. Any artefact produced by someone deep in context benefits from a reader with none. Marketing copy written by the person who wrote the brief. Image prompts written by the person who has been staring at the reference board for a week. A shot list assembled by the director who already knows what the scene is supposed to feel like.
The failure mode is identical in all of them: the artefact is judged against the author's intent rather than against what it communicates. A goldfish reviewer with only the artefact can tell you which of those two things it received.
That is also why the plugin's drift report is more interesting than its review command. Drift is the accumulated cost of context. The document said one thing, the code does another, and everyone who was in the room knows why, so nobody updates the document. Six months later the document is a liability. An agent that compares the two without knowing the reason is the only kind of reviewer that will flag it.
The cost of reviewing everything
There is a failure mode on the other side of this, and EGM's own structure hints at it.
Two agents mean two passes, and a review loop that runs on every draft will happily generate findings forever. Some of those findings will be real. Many will be a fresh reviewer rediscovering a constraint you already accepted for reasons it cannot see, which is the price of withholding context. You get blindness in exchange for independence, and independence is only worth paying for when the stakes justify a second opinion.
The plugin's framing handles this reasonably well. The goldfish reviews a specification before implementation, not a commit after every change. The drift report runs against a folder of design documents, not against the live repository on a hook. Those are the two moments where a forgotten-context opinion pays for itself: before you commit to a design, and after enough time has passed that the document and the code have separated.
The anti-pattern is treating EGM as a linter. A linter runs on every save because it is cheap and deterministic. A goldfish reviewer is neither, and a team that runs it on every save will start ignoring it within a week, which puts you back where you started, except now with a plugin installed.
A note on the star count
18,600 stars in a week is not a measure of quality. Google's logo is doing some of the work, and plugin repositories in this category have been accumulating stars at a rate that suggests a lot of people are bookmarking rather than installing.
What is worth taking from EGM is narrower and more durable than its popularity: the split between the agent that holds context and the agent that is denied it. That is a structural idea you can implement in an afternoon with two prompts and a file, whether or not you ever install the plugin. The hard part was never the code. It was admitting that the person closest to the work is the worst person to review it.
Related articles
Google replaced the model behind Search, and the number that moved was a shell benchmark
Gemini 4 Hydrogen became the default on 12 November, taking over from Gemini 3.8 Flash across Search, the Gemini app and the developer API. It runs about 2.5 times faster, scores 57% on Terminal-Bench 4.1 where its predecessor scored 18.2%, and costs $1.50 in and $7.50 out per million tokens.
Three ways a coding agent wastes money, measured on 30 repositories
TokenEconomizer watches what an agent did in a repository you work in often, then stops it from rediscovering the same build command, loading tools it never uses and looping on one error. Across five agent and model combinations it cut mean cost by 18.4% to 30.9%, cheaper in all 30 repository cells, with Pass@1 up in four of five.
earshot gives coding agents 922 sounds, and every one is rendered from code
An agent can write a game, an interface or a film and never hear any of it. earshot answers requests like "three variations of a footstep on gravel" from a catalogue of 922 designs built from public-domain recordings, with no audio files in the repository at all. It runs on CPU, as a CLI, a Python package, an MCP server and a Claude Code plugin.
The agents stealing API keys were not sophisticated. The gateways were open.
Amazon's decoy servers spent October posing as inference APIs and logged agents built on ARTEX pulling provider keys out of LiteLLM, Dify, Langflow and Open WebUI. The way in was a default master key and a half-finished install. ARTEX itself ships with budgets, scope gates and evidence for every finding.