richmondalake/memorizz/memorizz-agents

MemoRizz agents and harnesses

Saved agents

• memorizz_list_agents lists the agents this server exposes; memorizz_get_agent

and memorizz_inspect_agent describe one (instructions, memory, tools, delegates).

• memorizz_execute_agent runs one turn of an agent with a message and returns

its answer. It is off unless the user enabled it (memorizz config set MEMORIZZ_MCP_SERVER_ALLOW_AGENT_EXECUTION true); if the tool says it isn't allowed, tell the user that command.

• An agent's turn uses its own model and may cost money; say which agent you

are asking and why. Quote or summarise its answer and say it came from that agent.

Other harnesses

With harness execution enabled (memorizz config set MEMORIZZ_MCP_SERVER_ALLOW_HARNESS_EXECUTION true and MEMORIZZ_MCP_SERVER_HARNESS_WORKSPACE_ROOTS set to the allowed folders), memorizz_start_harness_run runs a bounded task on Claude Code, pi, Hermes, OpenHands, Codex or a saved MemAgent, in a project folder or a fresh scratch folder. Edits, web access and model-chosen verification commands wait for the user's approval in MemoRizz. Poll memorizz_get_harness_run for the result, and memorizz_cancel_harness_run to stop it. Never start the agent you are running in (it would run inside itself).

memorizz_start_harness_comparison gives one read-only question to several harnesses at once (each on its own model with harness_models), which is useful for a second opinion on a design or a bug.

A saved agent's own memory

• memorizz_list_conversations (agent_id) and memorizz_get_conversation

read a saved agent's conversations.

• memorizz_compact_conversation summarizes and compacts one of them, and

memorizz_compile_memory turns its continual-learning events into recall material. Both change the agent's memory, so run them only when the user asks.

• memorizz_query_traces (when the user enabled it with

memorizz config set MEMORIZZ_MCP_SERVER_ALLOW_TRACE_QUERIES true) shows what an agent did on a turn: model calls, tools and memory reads.