Mnemon

How it works

The /add-mnemon skill installs mnemon into the agent container image and wires it into the container lifecycle. Once applied, agents recall relevant memory before they respond and store new insights after each turn — so context carries across separate sessions instead of being lost when a conversation ends.

Mnemon works through Claude Code hooks. On each container start, the entrypoint runs mnemon setup --target claude-code, which registers hooks in /home/node/.claude/settings.json. Those hooks surface relevant memory ahead of the agent’s response and write back what it learns afterward. Because the hooks fire only under --target claude-code, this skill applies to agent groups running the default Claude provider — a group set to a different provider (for example "provider": "opencode") spawns its own process and never invokes the claude CLI, so the hooks don’t run for it.

Memory is stored at MNEMON_DATA_DIR=/home/node/.claude/mnemon, which maps to the per-agent-group .claude/ mount on the host. That location survives container restarts, so an agent’s accumulated memory is durable rather than ephemeral.

Setup

The skill runs in three phases:

1. Pre-flight — checks whether the Dockerfile already contains the mnemon layer, confirms the target groups run the Claude provider, and looks up the latest mnemon release to pin as the build version.
2. Apply changes — inserts the mnemon install block into container/Dockerfile, adds the idempotent mnemon setup call to container/entrypoint.sh (right after set -e, before stdin is captured), copies the Dockerfile and entrypoint structural tests into src/, then rebuilds and smoke-tests the image.
3. Restart and verify — restarts the service, confirms mnemon setup ran in the container logs, and inspects the registered hooks inside the running container.

Every step is idempotent. If the skill has already been applied, re-running it simply skips work that is already in place.

To confirm the wiring after a container starts:

docker exec $(docker ps --filter name=nanoclaw-v2 --format '{{.Names}}' | head -1) \
  cat /home/node/.claude/settings.json | grep -A5 mnemon

Tips

• The install layer pins a MNEMON_VERSION build arg. On an upgrade, the bundled mnemon-dockerfile.test.ts and mnemon-entrypoint.test.ts turn red if the install layer or the entrypoint wiring is dropped, so keep them in your test tree.
• To reset an agent’s memory, stop its container and delete the mnemon/ subdirectory from the group’s host .claude/ path. You can find that path with docker inspect on the running container’s mounts.
• If you see mnemon: command not found in the container, the image wasn’t rebuilt after the Dockerfile change — run ./container/build.sh and restart.
• To test recall, have a conversation, start a fresh session, and reference something from the earlier one. Mnemon should surface the relevant context without you restating it.