Editorial illustration for Google AI Releases Elephant-Goldfish Plugin; Issue #237 Reports AGENTS.md Leak
Google's Elephant-Goldfish Plugin Fixes AI Agent Memory
Google AI Releases Elephant-Goldfish Plugin; Issue #237 Reports AGENTS.md Leak
Google's open source team put out a new plugin this week called the Elephant-Goldfish Model, and the name alone tells you what problem it's solving. AI coding agents forget what they just agreed to, then confidently build the wrong thing anyway. EGM forces a different order of operations: the agent writes a specification first, then fresh subagents with zero conversation history get thrown at that spec to find holes before a single line of code gets written. The idea traces back to an essay by Dave Rensin, and Google says thousands of its own engineers already run it internally.
It's not a model. There are no weights to download. EGM is Apache 2.0 licensed and ships as a set of skills, reviewer personas, and document standards that sit on top of whatever LLM your harness already has access to, whether that's Codex, Claude Code, Claude Cowork, or Antigravity. The README is upfront that this isn't an officially supported Google product, which matters once you start looking at how the early test runs actually went across different harnesses, including one flagged in Issue #237 involving AGENTS.md leaking where it shouldn't.
Google has open-sourced the Elephant-Goldfish Model (EGM) plugin, a design-and-review layer for AI coding agents. The Elephant-Goldfish Model plugin makes an agent write a specification before it writes code. Fresh subagents with no conversation history then attack that spec.
Why this matters
EGM is a bet that the fix for bad AI-written code isn't a smarter model, it's better process: write the spec, let fresh reviewers with no memory of the conversation pick it apart, then code. Google saying thousands of engineers already run this internally gives the approach real weight, not just a GitHub demo. For teams building with AI agents, the appeal is obvious: catching a wrong assumption in a Markdown file costs a few minutes; catching it in a pull request after the code ships costs a lot more.
But Issue #237, where AGENTS.md leaked into Codex reviewers, is the detail worth watching. The whole design hinges on "fresh subagents with no conversation history" staying isolated from each other. If context bleeds across reviewers, the independence that makes this work disappears, and you're back to one biased opinion wearing nine personas.
Add in the token cost of running 5 review stages and 9 personas per spec, and EGM looks less like a free lunch and more like a trade: slower, pricier reviews in exchange for fewer surprises later. Worth testing before trusting on anything production-critical.
Related Reading
Browse all Google news →
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Axiom Math raises USD 200M Series A led by Menlo to keep AI slop out of math
Related Topic Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI
Related Topic NVIDIA and Google Cloud let developers scale AI from prototype to production
Market Trends
From Archives Meta’s Muse AI creates custom app spaces by letting users describe them
From Archives PrimusClaw Open-Source Runtime Manages AMD Instinct GPU Clusters
From Archives LangSmith BYOC Expands to AWS, Azure, and Google Cloud
Research & Benchmarks
From Archives Google's Gemini Omni 1.2 Flash Extends AI Video to 15 Seconds, Cuts Prices 30%
LLMs & Generative AI
Common Questions Answered
How does the Elephant-Goldfish Model plugin prevent AI coding agents from building the wrong thing?
The EGM plugin forces AI agents to write a detailed specification first before writing any code. Fresh subagents with zero conversation history then review that specification to identify holes and flaws, ensuring assumptions are validated before development begins. This process-based approach prevents agents from confidently implementing incorrect solutions based on forgotten context.
What is the key difference between EGM's approach and traditional AI code generation?
Instead of relying on a smarter model to generate correct code directly, EGM implements a design-and-review layer that separates specification writing from code implementation. The use of fresh subagents with no conversation memory to review specs eliminates the problem of agents forgetting previous agreements and ensures independent validation of requirements.
Why does Google's internal use of this process give the Elephant-Goldfish Model credibility?
Google reports that thousands of engineers already run this specification-first, fresh-reviewer process internally, indicating it's a proven methodology rather than just a theoretical GitHub demo. This real-world validation at scale demonstrates that the approach effectively catches wrong assumptions early in the development cycle when fixes are inexpensive.
What cost savings does catching errors in the specification phase provide compared to code review?
According to the article, catching a wrong assumption in a Markdown specification file takes only a few minutes to fix, whereas catching the same error in a pull request after code has been written is significantly more expensive and time-consuming. This early detection approach dramatically reduces the cost of correcting fundamental misunderstandings about requirements.
Further Reading
- Why AI coding agents forget: memory, handoffs, and latent context - AWS Builder
- How Does a Team Share Context with AI Coding Agents? - AQ
- File-based agent memory - Tom Rochette
- Managing Agentic Memory is a New Job for Specialized Memory Agents - HackerNoon
- Coding Agent Memory Compared 2026: What Survives a Session - IA Decider
