Agent Memory Trellis— plugin

<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"> The fundamental 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, and your own, by growing your project along clean documentation lattices. Give agents only new instructions, and watch them pick up where you left off, as instant experts in your repository.

The skills maintained here bear the brunt of the repo-hygiene / anti-entropy / anti-bloat work needed to support this. They are — in this spirit — deliberately minimal, project-agnostic, and respectful of existing conventions. Feel free to use them as-is, or build them out to your preferred specs.

Contributions are welcome — see guidelines.

The skills

#### 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.

Rationale

The skills address LLM tendencies that are likely to persist through jumps in intelligence and context capacity. Our fundamental assumptions:

1. LLMs are inferencing machines. 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" alone, they are liable to losing the plot to an inundation of their own making.
2. Context should be concise. We experience this as a context-size tipping point, beyond which model performance degrades. They begin to struggle with prioritization, or need frequent reminders of earlier details or guidelines. Just like you and me, models need some free overhead space to mull ideas and grow into.
3. Context should be maximized. Nonetheless, agents do their best work if they have maximal access to relevant context for the specific problem they are actively working on.
4. We should prioritize clean re-entrancy. To support this, it helps to give agents headroom for new tasks, leaving a clear path for retrieving old, relevant details and resolving inconsistencies. In the limit, the clarity / navigability of this path can be measured by new agents seeing the repo for the first time.
5. We should use progressive disclosure. To avoid inundating context right at our entry points, we should aim to store only essential information and indices, pointing to progressively more complete artifacts of existing code, docs, discussion etc.
6. We should aim to self-right. Finally, we must accept that new work will produce new noise and entropy, and even agents tasked with cleanup will miss things. As long as our systems are self-righting — front-and-center docs nudge agents back towards order — our projects can stay upright even through rapid growth.

Information privacy

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.

Install

Two paths — depending on your goals:

Consume (stay current)

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-trellis

In 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.

Own (customize)

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.

License

MIT