What is claude-mem, and how to use it in Claude Code

The moment it pays off

You spend a session working out why a build fails. You find it. The session ends.

Tomorrow you open Claude Code and the agent knows none of it. Not the dead ends you ruled out, not the file that turned out to matter, not the fix. You explain it again, and the explaining is the expensive part.

claude-mem is built for that gap. It watches what happens during a session, compresses it, and puts the relevant part back in front of the agent next time.

What is actually inside

Nineteen skills and one command. The full set, not a sample:

babysit · cloud-sync · design-is · do · how-it-works · knowledge-agent · learn-codebase · make-plan · mem-search · mode-creator · oh-my-issues · pathfinder · smart-explore · standup · timeline-report · version-bump · weekly-digests · what-the · wowerpoint

The command is anti-pattern-czar.

Read the names and the shape of the thing shows up. Some are about getting context in (learn-codebase, smart-explore). Some are about getting it back out (mem-search, timeline-report, weekly-digests, standup). Others are working habits the author wanted to keep across sessions rather than retype (make-plan, babysit, do).

That is a description of the names, not a claim about the code. For what each one does, read it at the source: every skill on the listing links to its own file.

Install it

/plugin marketplace add thedotmack/claude-mem
/plugin install claude-mem@thedotmack

The first line adds the marketplace once. The second installs the plugin.

How invocation actually fires

Every other plugin directory stops at the command above. This is the part that decides whether any of it matters, and the answer is different for each kind of component.

A skill can fire itself. The model reads the skill's own description and invokes it when your prompt matches, without you naming it. You can also call one by name, and each is available as a slash command. Nineteen of the twenty components here are skills, so most of this plugin is the kind that can act on intent rather than on instruction.

A command never fires itself. anti-pattern-czar runs when you type it. That is the whole contract, and it is a feature: some things should only happen because you asked.

Hooks run on events, not on intent. They fire when something happens in the session rather than when your prompt suggests they should. claude-mem ships hooks, which is consistent with a plugin whose job is to notice things while you work.

The distinction is worth holding on to, because "I installed it and nothing happened" almost always means the component was a command and nobody typed it, or a skill whose description did not match how the prompt was phrased.

Who it is for

People working on the same thing across many sessions. A long refactor, a system you return to weekly, a codebase you are still learning. The value comes from accumulation, so a plugin like this is worth least on day one and most in month three.

If your work is a series of unrelated one-off sessions, memory across them buys you less.

One thing it does not do

It does not make your other skills fire. Context loss and skill invocation are different problems that get confused because both feel like the agent forgetting. claude-mem addresses the first. Whether an installed skill actually gets invoked at the right moment is the second, and it is answered by the routing layer on an auto-invoked listing.

Worth being clear about, because a memory plugin is the closest thing in the catalog to the compaction problem, and it would be easy to oversell it into solving the other one.

Read it before you install it

The listing carries the full component list, every skill linked to its source, and the repository behind it. Nothing here is a substitute for reading the file you are about to hand your agent.