Memctl: Versioned memory for your coding agent (Claude.md / AGENTS.md)

What happened

- Tool name: memctl (by Elton Cherrington) - Repository: EltonCherrington/memctl on GitHub - Description: "Version, share, restore AI-agent memory files (CLAUDE.md/AGENTS.md) from one central store. Zero-dependency CLI." - Works with: Claude Code, Cursor, Copilot, and any tool that reads those files - Not yet on npm (publish pending); install from source - Commands include: memctl init --template=node, capture/save, restore, export pack - Optional env var relocates the store (dotfiles, synced folder, own server) - License: MIT, built in one session by an autonomous agent - Starter pack: CLAUDE.md, stack templates, cheat sheet - Premium starter pack: $5 USDC on Base, on-chain verified unlock - Wallet: 0x648bAa08901f1bEAB002Af57f1375F80Ec4F4893 - Pay-what-you-want donations: $3/$5/$10 - HN post: 1 point, 0 comments - Key insight: "In a fresh clone you get a fresh, forgetful agent" - Files are "copied (never mutated) to and from" the store - Built-in starter pack included - Can export a pack to share with a teammate

Now I'll write the article: ---

Elton Cherrington has released memctl, a zero-dependency command-line tool that versions, shares, and restores AI-agent memory files — specifically CLAUDE.md and AGENTS.md — from one central store. The project, published on GitHub under the MIT licence and built in a single unattended session by an autonomous agent, treats those configuration files the way a package manager treats dependencies: as explicit, restorable project artifacts rather than ephemeral scratch files that get wiped on every fresh clone.

How it works

This piece explains what memctl does, how it stores and moves memory files between projects and machines, why the problem it addresses has become more acute as Claude Code, Cursor, and Copilot have proliferated, who is most likely to reach for it, and what a developer should verify before trusting it in a production workflow. It is aimed at builders who already use at least one AI coding agent and have noticed inconsistent behavior across clones or teammates.

Cherrington shipped memctl as a public GitHub repository at EltonCherrington/memctl. The tool provides a CLI for saving a project's CLAUDE.md or AGENTS.md into a central local store, restoring that saved copy into any other directory, and exporting the result as a shareable pack file. A built-in starter pack ships with the repository and includes a CLAUDE.md, stack templates, and a cheat sheet. A premium starter pack — described as a purpose-built template wired for a specific stack, with Claude Code-powered install, test, and build helpers — is available separately for $5 USDC on Base, unlocked on-chain at a buy page hosted at raw.githack.com/EltonCherrington/memctl-shop/main/site/index.html. Donations to the author's Ethereum address (0x648bAa08901f1bEAB002Af57f1375F80Ec4F4893) are accepted at pay-what-you-want tiers of $3, $5, and $10. The package is not yet on npm; at the time of publication, installation requires cloning the repository from GitHub.

The announcement surfaced on Hacker News with 1 point and 0 comments, which means there is no community feedback yet to draw on. The project's own documentation and README are the primary source of ground truth for any developer evaluating it now.

Why it matters

When a developer runs the save or capture command in a project directory, memctl copies the current CLAUDE.md or AGENTS.md into a central store rather than modifying or symlinking the originals; the repository description explicitly states files are "copied (never mutated) to and from" the store. A restore command writes the saved copy back into a target directory, making the same memory file available in a fresh clone or on a new machine without manually hunting for a prior version. The init command — shown in the docs as memctl init --template=node — scaffolds a new memory file from a template with the project name pre-filled. An export command produces a .memctl.pack file that a developer can hand to a teammate.

The location of the central store can be overridden with an environment variable, so a team can point it at a dotfiles folder, a synced cloud directory, or a self-hosted server. Because the tool has zero external dependencies and is installed from source, there is no npm registry, no token, and no network call during normal use. The only external component is the optional on-chain payment flow for the premium starter pack, which uses Base (an Ethereum layer-2) and does not require an account.

The core problem memctl addresses is agent amnesia across repository boundaries. As Cherrington states in the README: in a fresh clone you get a fresh, forgetful agent. Memory files like CLAUDE.md and AGENTS.md contain the rules, context, and constraints that shape how an AI agent behaves in a project — its allowed commands, its code-style preferences, and any domain knowledge the team has encoded over time. When those files are missing, the agent reverts to generic behavior, and engineers spend time re-establishing context that already existed somewhere else.

Who is affected

There is currently no standardized tooling for treating these files the way teams treat dependency lock files or Docker base images. Memctl proposes a simple answer: give memory files the same copy-on-save versioning discipline that developers already apply to build artifacts. MIT licensing and zero dependencies keep the adoption cost low. The fact that the tool was itself built by an autonomous agent in one session and published as a working product is also a concrete demonstration of the exact workflow it supports — using Claude Code or a similar tool to produce shippable software end to end.

The immediate audience is any developer using Claude Code, Cursor, GitHub Copilot, or any other AI coding tool that reads CLAUDE.md or AGENTS.md from a repository root. Teams that share a codebase across multiple contributors will benefit most from the pack-export feature, which lets one engineer's carefully crafted memory file be distributed to others without relying on git history or manual copying. Freelancers and solo developers who switch between many projects and want to carry context from one engagement to the next are also a natural fit.

Platform authors who ship their own AGENTS.md conventions — such as monorepo maintainers or open-source project leads — could use memctl to distribute canonical starting configurations the same way they distribute Prettier or ESLint configs today. The $5 Base payment for the premium starter pack is low enough to be a non-issue for professional engineers but does introduce an on-chain dependency that enterprise security teams may need to review before permitting it in a procurement process.

What to watch next

The most important near-term signal will be whether the npm publication completes. Installation from source creates friction and means users cannot rely on a registry-pinned version, which matters for reproducible CI environments. A builder evaluating memctl now should verify that the copy-never-mutate guarantee holds for their file watcher or hot-reload setup — specifically that no background process in Claude Code or Cursor is watching the central store path and re-reading from it as though it were the project's own directory.

Cherrington links to a prose field report titled "72 hours unattended — the memory bugs behind this tool" at raw.githack.com/EltonCherrington/memctl-shop/main/site/essay.html, which is worth reading for practical failure modes the author encountered during the build session. Watch the GitHub repository's issues and the eltoncherrington.github.io/memctl site for documentation improvements, a conflict-resolution story for divergent memory files across branches, and any move toward a registry release that would let teams pin a specific memctl version in their own toolchains.

Full article saved to [memctl-article.md](file:///home/ubuntu/.gemini/antigravity-cli/brain/d5ed5091-8455-4214-83f6-815364d8e444/memctl-article.md).