Why mehmory

The problem

Every Claude Code or Codex CLI session starts from zero.

You spend the first twenty minutes re-explaining the architecture, the deploy story, why that one module is weird, and the decision you already made twice. Then /clear lands, or the context window compacts, and the session that finally understood your project is gone. Tomorrow you do it again.

The usual answers ask for a lot of trust. A vector database you can't read. A hosted service holding your codebase's institutional knowledge. An MCP server between you and your own notes. Retrieval quality nobody measures, in a store nobody can open.

The answer

Your project's memory is a directory of markdown files in a git repo at ~/.mehmory.

That's the whole storage layer. A Claude Code or Codex CLI session reads it and writes it with ordinary file operations. You can cat it, grep it, edit it in your editor, git diff it, and revert a bad write with git revert. When memory is a text file, "what does it think it knows?" is answerable in one command instead of being an act of faith.

mehmory search "what did we decide about auth"

What "meh" buys you

The name is a promise about scope. mehmory is not trying to be a cognitive architecture. It does one boring job well, and the boring choices are the point:

Nothing to run. No server, no daemon, no MCP process, no API key, no account. A CLI, a handful of hooks, and a folder.

Nothing to lose. Your memory is in your filesystem, under your git, on your machine. Uninstalling the plugin doesn't delete it — and mehmory is explicit that those are two different operations, with a documented export and restore path for the one that does.

Nothing to break your session. Every hook fails open. If a hook errors, times out, or finds a corrupt store, the session continues as if mehmory weren't installed. Memory tooling that can take down your editor isn't a productivity gain.

What it actually does

Captures without asking. Deterministic hooks distill your sessions into an inbox as you work — no model call, no token spend, no interruption. They skip the harness's own noise and turns too short to be a fact, so the inbox is what you said, not a transcript. mehmory onboard backfills the same inbox from the Claude Code transcripts you already have, so a project with history doesn't start empty.

Integrates with judgment. Sorting a fact into the right page is a reading-comprehension problem, so a model does it. /mehmory:integrate reads the inbox and the schema, decides where each fact belongs, writes the pages, updates the index, and commits.

Ages gracefully. Memory that never forgets is memory that drowns you in last quarter's decisions. Aged pages get demoted and flagged, never silently dropped — you can always see what was down-weighted and why.

Stays inside a budget. Two things load before you type a word: the injected context frame and the plugin's skill descriptions. Both are capped, and both caps are enforced by tests. Raising one is a commit that says why, not a surprise you discover in your token bill.

Redacts before it writes. The secret filter runs on the way into the store, and its real limits are documented rather than oversold.

Honest limits

• Retrieval is keyword matching over one scan. No index, no embeddings. Fast, greppable, debuggable — and genuinely worse at paraphrase than a vector store. The paraphrase gap is measured and published in the golden set rather than hidden.
• The first session after onboarding sees a stub. Onboarding seeds raw material; the wiki gets built by the first /mehmory:integrate. The session where "it already knows my project" comes true is the second one. Anything claiming otherwise is overselling it.
• Integration is in-session work. It costs tokens, because judgment costs tokens.
• onboard's transcript backfill is Claude Code only. It mines ~/.claude/projects/*/; there's no equivalent transcript history to mine on Codex, so a Codex project starts from the hooks capturing forward, same as a fresh Claude Code project would.
• Codex's hook-trust behavior on first run is unverified. Codex CLI can require an explicit trust decision before running a freshly written hook; whether that gates mehmory's first Codex invocation is not something this project could measure without a real user's configuration. See the troubleshooting doc.

Convinced enough? Quickstart takes about five minutes. Curious how it's built? How it works.