Skip to content

MemoRizz v0.17.26

Latest

Choose a tag to compare

@github-actions github-actions released this 16 Nov 15:18

Added

  • Voice for agents. MemAgent(voice=True) gives an agent a voice through
    LiveKit; voice_config picks ElevenLabs, Vapi or OpenAI instead, or names
    the voice and models. An agent can speak a reply (speak), transcribe a
    recording (listen), take a whole spoken turn (run_voice) and, with
    LiveKit or Vapi, hold a live call (start_voice_session). Voice settings
    are saved with the agent; keys never are, and are set in the environment or
    under Settings → Voice. The agent form has a Voice section,
    memorizz agents create and update take --voice, --voice-provider and
    --voice-name, the MCP server's create and update agent tools take the
    same three, and pip install "memorizz[voice]" adds LiveKit's packages.
    Guide: Voice.
  • Tools an environment lends. agent.enter_environment(name, tools) adds
    tools to an agent for as long as it is in an environment, and
    leave_environment(name) takes them away. They are offered on every turn
    and are never written to the saved agent or to toolbox memory, so a
    simulation or a session can give an agent abilities without leaving
    anything behind.
  • A rejection the model hears about. agent.resume_rejection(proposal_id)
    answers the paused tool call with the refusal and the approver's reason, and
    the model continues the run: it can say so or do something else. Before, a
    rejected proposal left the run paused and the host had to write the next
    message itself. Guide: Tools and approvals.
  • MCP tools before the first turn. agent.discover_mcp_tools() lists the
    configured servers and registers each tool as server__tool, so the first
    turn no longer spends steps on mcp_list_tools. only_direct_tools=True
    also removes the general MCP tools. Calling with_mcp_servers again now
    registers tools listed since the first call.

Changed

  • A paused turn stores a waiting note, not the proposal. The reply kept in
    conversation memory for a turn that stopped for approval is one line with
    the proposal id. When the proposal is resumed or rejected the line is
    replaced by what happened: the model's reply, or that the call ran, failed
    or was rejected.
  • A call through mcp_call_tool is recorded under the tool's own name
    (calendar__create_event) in the approval proposal, the workflow step and
    the tool log, when that tool is registered. A proposal for such a call now
    has that tool_name and the tool's own arguments.
  • A refine cycle counts what it stored for review as proposed_
    (proposed_tool_revisions, proposed_standing_rules,
    proposed_persona_additions). planned_ is kept for a dry run. The tool
    review's note names MCP tools as such (they are not reviewed), and the
    persona job's note is labelled "Persona addition", so it no longer reads as
    a verdict on a standing rule proposed in the same review.
  • Tool learning may name a constant the tool expects. A usage note can
    say "set source_thread to owner" when that plain word put a failed call
    right and the tool's error names it, or it put two failed calls right. Ids,
    names, amounts and anything with a digit or a space are still refused.

Fixed

  • A decided approval was recalled as still waiting. The pause payload
    stayed in conversation memory as the turn's reply and the resumed reply was
    never stored, so later turns, in the same thread or another, said a booked
    event was "still waiting for your approval".
  • Calls held back by an approval were never made. When a model asked for
    several guarded calls at once, the others were answered "waiting for human
    approval" and that text was still what the model read after the resume, so
    it reported them as held. They now say the earlier call is decided and
    that they may be made.
  • A request was stored as a fact, and one preference under several names.
    The stated-facts job no longer files "Add a task: ..." as user.task, and
    a statement memory already holds under one attribute is not added again
    under another name, in the same cycle or a later one.
  • apply_persona_rule says that no standing rule is waiting, or lists the
    ids that are, instead of "No standing rule is waiting under that id".