i almost switched my claude code memory to a shinier tool. then it hit me: i’d already solved this years ago, and the proof is literally the post youre reading right now.
how my memory actually works
my agent’s long-term memory is boring on purpose: plain markdown files on my disk, plus this blog. i set it up this way ages ago, before claude code shipped any of its built-in memory features.
then claude-mem showed up. it captures everything the agent does, compresses it with AI, and auto-injects it into the next session. felt like an upgrade, so i almost ran it on top.
the 2 questions that stopped me
q1: where does the “compression” run? it says “local database”, but the step that turns raw stuff into memories calls a remote model. so every tool call ships my file contents out to be summarised. normal repo, fine. my /Personal folder (credentials, contracts, drafts), not fine. and its opt-out, which is a tripwire.
q2: rebuildable index, or the only copy? episodic-memory (what i use) just indexes logs already on my disk, delete it and nothing’s lost. claude-mem captures into its own private db, delete it and its gone.
the thing i noticed: both answers are stuff plain markdown already nails. its local, its mine, i can read it with my own eyes, and nothing leaves unless i choose to write it down.
so what did i do
stayed put. plain markdown + episodic-memory. claude-mem isnt bad, it just failed both questions for my setup and duplicates what i already have.
but the real takeaway is bigger than the tool.