Agent Memory Trellis — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited Agent Memory Trellis (Plugin) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
<p align="center"> <img src="assets/trellis.png" width="200" alt="A green vine growing up a wooden trellis"> </p>
<h1 align="center">agent-memory-trellis</h1>
<p align="center"> A minimal skill kit for agentic engineering with persisted memory. </p>
<p align="center"> <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-blue.svg" alt="MIT License"></a> <img src="https://img.shields.io/badge/Claude%20Code-plugin-d97757.svg" alt="Claude Code plugin"> </p>
Play to your agents' strengths by growing your project along clean documentation lattices. Give new agents only new instructions. Watch them pick up work easily, while exercising caution around their weaknesses... and proactively helping you keep your repo clean.
The hope is that these skills free you up to work faster and more confidently, by automatically directing agent attention regularly to repo hygiene, entropy, and bloat.
Some reasons why this reminding seems needed are offered below (and might also be useful context). The skills themselves, though — in the spirit of anti-bloat — are intentionally kept small, agnostic, and respectful of existing project conventions.
Feel free to use them as-is, or customize to your needs. Contributions welcome — see guidelines.
| Skill | Responsible for | Invoke when |
|---|---|---|
doc-library | Structure — what docs exist and where they live | Creating docs for the first time, or seeking guidance on doc responsibilities |
grill-me | Planning — make details explicit, and interrogate human <> agent gaps | Planning new features (an extension of the famous Matt Pocock skill, with nuanced doc guidance) |
handoff | Continuity — session state, saved and picked up | Ending a working session; or starting the next one |
doc-audit | Verification — do the docs match the code | Major code changes or architecture planning have occurred, or if docs feel stale |
ubiquitous-language | Vocabulary — a canonical domain glossary | Shorthanding complex ideas, or resolving ambiguous terminology |
#### Note on multi-agent runs:
Supported out of the box: under `handoff`, parallel agents each write their own scoped snapshot — without locks — and a coordinator rolls them up. Single-agent use ignores this entirely.
Our skills address LLM tendencies that seem likely to be sticky through technological advances:
Modern LLMs, built on the transformer model, are fundamentally better at pattern-matching and distilling information, than they are at foreseeing the long-term consequences of their own actions. Set loose on "vibes", they are liable to lose the plot to inundations of their own creation.
There seems to be a context-size tipping point, unique to each model, beyond which performance degrades. Models begin to struggle with prioritization, or need frequent reminders of earlier guidance. Just like us, models benefit from having free overhead space to grow into.
Despite their finite attention capacity, agents do their best work if they have maximal access to relevant context for the specific problem they're working on.
To optimize for both context conciseness and depth, information architectures must be easily navigable. A new agent should always have a clear path for retrieving relevant details and understanding the full scope of a task.
To avoid inundating context right at our entry points, we should aim to store only essential information and indices there, pointing to progressively more complete information (existing code, docs, discussion etc.)
Finally, agents must clean up after themselves. On an individual basis, we should accept that this will be imperfect — agents will miss things and make mistakes. As long as the overall architecture is self-righting — the entry-points state clear expectations and nudge things back toward order — then our project should stay upright through growth.
For public repos, a gitignored docs/private/ tier may be kept with its own INDEX; the public INDEX points to it in one line, and public docs reference it in plain text so they stand alone in any clone. doc-library defines the possible split; doc-audit polices it to prevent leaks and broken boundary links.
Two paths — depending on your goals:
Install as a Claude Code plugin and get auto-updatable, namespaced skills (/trellis:*).
In the terminal CLI:
/plugin marketplace add s-frantz/agent-memory-trellis
/plugin install trellis@agent-memory-trellisIn the VS Code extension, the /plugin command isn't available — open the `/plugins` panel instead, add the s-frantz/agent-memory-trellis marketplace under the Marketplaces tab, then install trellis. Both surfaces share the same ~/.claude/ config, so a plugin added on one shows up on the other.
Installed skills live outside your repo, in Claude Code's plugin cache.
Copy the skill folders into your project — bare /skill-name invocation, helpful if you want to customize them:
# Claude Code, per-project
cp -r skills/* {your-repo}/.claude/skills/
# Claude Code, global (all projects)
cp -r skills/* ~/.claude/skills/Other agent harnesses that accept markdown instructions can use each skills/*/SKILL.md as-is; only the /slash-command invocation is Claude Code-specific.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.