mehmoryYour project's memory, in markdown

A deliberately "meh"-tier improvement in memory and continuity for Claude Code and Codex CLI, built from boring parts. Git-backed markdown at ~/.mehmory. No embeddings, no MCP server, no cloud.

Quickstart

Why mehmory

GitHub

Boring parts, on purpose

A directory of markdown in a git repo. Read it in your editor, grep it, diff it, revert it. Nothing in the storage layer is a black box you have to trust.

Capture is deterministic

Hooks on SessionStart, UserPromptSubmit, Stop, PreCompact and SessionEnd do the capturing. No LLM call in the hot path, no token spend, and every hook fails open — a broken memory layer never breaks your session.

Judgment is model-driven

Deciding what a fact means and where it belongs is the model's job, not a regex's. /mehmory:integrate reads the inbox, writes pages, updates the index and commits.

Budgeted, not bloated

The always-on context frame is capped at 800 tokens and the cap is asserted by a test. Memory that quietly eats your context window is a regression, not a feature.

Retrieval you can measure

A golden query set reports Recall@1 and Recall@3 on every scoring change, and reports the paraphrase split separately — the measured size of what keyword matching cannot do.

Local-first and deletable

Secrets are redacted on the way in. purge removes pages, a session's un-integrated captures, projects or everything, with an export first. mehmory never rewrites your git history.