MemoryCustodian

What is MemoryCustodian?

MemoryCustodian is an open-source, local-first tool that gives coding agents—like Codex, Claude Code, and Gemini—durable project memory without requiring a hosted service or bloating every prompt. It stores decisions, constraints, rejected approaches, and project context as plain Markdown files directly in your repository. A manifest system loads only the memory relevant to each task, keeping agent sessions focused and efficient. MemoryCustodian is designed to be reviewed, versioned, shared, and deleted like code, making it a natural fit for any development workflow.

Who it's for

• Solo developers who want their coding agents to remember project context across sessions without manually pasting instructions into every prompt.
• Teams collaborating with AI agents who need shared, version-controlled project memory that everyone—human and agent alike—can access and update.
• Open-source maintainers who prefer local-first, self-hosted solutions and want to avoid vendor lock-in or recurring service costs for agent memory.

Key features

Plain Markdown storage

Memory is stored as standard Markdown files in your repository, so humans can review, edit, and version them just like code. No proprietary formats, no external databases—just files that work with your existing Git workflow.

Manifest-driven context loading

A manifest.md file routes the agent to only the memory files relevant to the current task. Instead of loading everything, the agent reads brief.md for project shape, decisions.md and constraints.md for planning, and do-not-use.md to avoid rejected paths—keeping context minimal and focused.

Cross-agent compatibility

MemoryCustodian works with Codex, Claude Code, Gemini, and other coding agents. A single init command creates bootstrap files for your chosen agent, and the same memory structure works across different tools and platforms.

Scoped and optional memory

Optional directories like rules/, profiles/, areas/, and archive/ stay out of the default context until they are explicitly relevant. This prevents memory bloat and ensures agents only load what they need for the task at hand.

What stands out

"Durable memory. Minimal context."

This one-liner captures MemoryCustodian's core philosophy: it gives agents lasting project knowledge without inflating every prompt. By storing memory as plain Markdown in the repo and loading only what's needed per task, it avoids the common trade-off between memory depth and prompt efficiency. The result is a system that feels invisible—agents remember what matters, and you never pay for context you don't use.

Worth checking out if…

You're tired of repeating the same project context to your coding agents every session, or you want a lightweight, open-source alternative to hosted memory services. MemoryCustodian is especially useful if you value local-first tools, version-controlled project knowledge, and the ability to review and curate what your agents remember. It's a practical fit for any team or solo developer using AI coding assistants who wants durable, minimal, and shareable project memory.