Stemma

Stemma gives coding agents notes on the production practices they miss, such as evaluation gates, safe reindexing, and access control. It follows the open Agent Skills standard, so it works in any harness that supports it. The agent opens only the notes a given decision needs. It has been submitted to Anthropic’s plugin directory and is in review.

Learning what AI memory is

When I first started learning about agentic systems, memory was a hot topic. There were dozens of terms I couldn’t follow, and I wanted to understand how AI actually remembers things, both within a conversation and across large amounts of context. I found that to understand agents, or anything in this space, I had to understand memory. So I asked Claude to build me a learning ladder for retrieval-augmented generation (RAG), worked through it one rung at a time, asking questions as I went, and built my first RAG project: 401(k) Plan Advisor. Then I read books, watched videos and took notes to formalize what I knew, and kept at it until even advanced RAG tutorials looked like naive RAG next to my notes. I built SEC Risk Analyzer, Campaign Intel and FDA Guidance Diff, each in less time than the last, and each more robust.

From notes to a skill

A few months later, I revisited my notes, because things change so fast, and did a full sweep to update them. Some ideas were deprecated; some that had seemed cutting-edge had fallen out of favor or been formalized, and a new frontier had opened. I needed to capture that in my notes and update my own mental models. But at the pace AI is moving, I also wanted my agents to use this knowledge base without my adding the context every time, and the easiest way to do that was a skill. So I packaged it up to use as I built more projects over large corpora, with these rigorous RAG processes on hand. I’m building FAA Directive Impact with it now.

Now I can keep updating the skill so it evolves over time: in its content, through periodic updates, and as a skill, using what a plugin can do. Use it when planning, even with frontier models; it stays current with the field. It’s open source, so reach out with questions or suggestions.

Does it work

Frontier models are good now, so I tested whether the notes still matter. An agent planned the path to production for one of my projects, SEC Risk Analyzer, three times with the skill and three times without, and each plan was checked for seven production practices. The skill was never tuned on this project, and the checks were written before any run. With the skill, Claude Opus met all 21 checks; without it, 5. The smaller Claude Haiku went from 5 to 16.

The models know RAG techniques, but that isn’t enough for systems like the ones I build, over retirement plans, SEC filings, FDA guidance, and FAA directives, where a wrong or unsupported answer has a cost. The skill supplies what’s missing: what’s good when, and how to actually build those systems. That gap between knowing the techniques and shipping a system people can rely on is what Stemma closes.

In Claude Code:

claude plugin marketplace add tanayvenkata/stemma
claude plugin install stemma@tanayv

Other agents, including Codex, Cursor and Gemini CLI, install it from GitHub.

The notes

Every note is my own synthesis, cited where a claim depends on a source and checked against primary sources. A script checks every link and reference on each change.

inventory source classes and risks
→ establish the simplest credible parser for each class
→ run deterministic completeness and integrity gates
→ inspect representative and high-risk samples
→ escalate failed pages or documents to specialized extraction
→ preserve provenance and unresolved failures
→ evaluate downstream evidence recovery