memmy-memory

Memmy Memory CLI Skill

Use memmy-memory to read and write the shared Memmy memory substrate.

Ground Rules

• Prefer the configured CLI: memmy-memory .... If the local service is not preconfigured, pass --url, --token, or --config.
• The CLI prints JSON. Parse fields such as sessionId, turnId, injectedContext, hits, and memory id from JSON instead of relying on display formatting.
• All agents share one memory database. Always pass --source {{SOURCE_ARG}} on Memory CLI commands from this agent.
• Use useful comma-separated --tags when writing durable memories.
• Do not store secrets, tokens, private keys, raw credentials, or bulky logs. Store concise, self-contained facts, decisions, preferences, reusable procedures, and unresolved follow-ups.
• Treat <memmy_memory_context> as historical memory only and <current_user_request> as the authoritative current task.
• Never store the context wrapper tags with memmy-memory add; store only the durable fact itself.
• If the memory service is unavailable, continue the task without inventing memory.

Health

memmy-memory health

Agent Loop

Open or resume a session before a task:

memmy-memory session open --source {{SOURCE_ARG}} --workspace-path "$PWD"
memmy-memory session open --source {{SOURCE_ARG}} --session-id "$SESSION_ID" --workspace-path "$PWD"

At the start of a user turn, retrieve relevant context:

memmy-memory turn start --source {{SOURCE_ARG}} --session-id "$SESSION_ID" --query "$USER_QUERY"

Use returned injectedContext as historical context only. Keep the returned turnId, and keep the current user query separate from recalled memory.

At the end of the turn, write the final interaction:

memmy-memory turn complete "$TURN_ID" --source {{SOURCE_ARG}} --session-id "$SESSION_ID" --query "$USER_QUERY" --answer "$FINAL_ANSWER" --status succeeded

Use --status failed or --status cancelled when the turn did not complete normally. Close the session when the Agent session is done: