NeMo Agent Toolkit Memanto Plugin

Long-term memory for NVIDIA NeMo Agent Toolkit workflows, backed by Memanto.

The plugin registers a memory backend with _type: memanto. It works with the toolkit's built-in memory tools (add_memory, get_memory, delete_memory) and with the automatic memory wrapper (auto_memory_agent).

How it stores memory

• One Memanto agent per user. Each user_id maps to its own agent, <agent_prefix>_<first 32 hex characters of sha256(user_id)>. Users are isolated by namespace, and user ids never appear in agent names.
• Explicit memories are stored as written. When a MemoryItem has memory text, as the add_memory tool usually produces, that text is stored with the item's tags.
• Conversations are distilled. When an item has only a conversation, as the automatic memory wrapper produces, Memanto extracts the durable facts, preferences and decisions in it and stores those as typed memories. Small talk stores nothing.
• Reads never create agents. Searching or deleting for a user with no memories returns nothing and creates nothing.

MemoryItem.metadata is not stored. Search results carry each memory's id, type, title, confidence, provenance, source and timestamps in metadata, and its relevance in similarity_score, where higher is more similar.

Install

uv add nemo-agent-toolkit-memanto

or:

pip install nemo-agent-toolkit-memanto

The package must be installed in the same Python environment as nvidia-nat-core. Confirm the toolkit discovers it:

nat info components -t memory -q memanto

Credentials

Set your Moorcheh API key, from console.moorcheh.ai:

export MOORCHEH_API_KEY=...

If it is not set, the key saved by memanto setup is used. The on-prem backend needs no key.

Configuration

memory:
  user_memory:
    _type: memanto

Automatic memory

The wrapper saves every user message, recalls relevant memories before the inner agent answers, and saves the agent's replies. The user identity comes from the toolkit runtime.

memory:
  user_memory:
    _type: memanto

llms:
  nim_llm:
    _type: nim
    model_name: meta/llama-3.3-70b-instruct

functions:
  current_datetime:
    _type: current_datetime
  my_react_agent:
    _type: react_agent
    llm_name: nim_llm
    tool_names: [current_datetime]

workflow:
  _type: auto_memory_agent
  description: Assistant that remembers each user
  inner_agent_name: my_react_agent
  memory_name: user_memory
  llm_name: nim_llm
  search_params:
    top_k: 5

Automatic memory needs the LangChain integration, nvidia-nat-langchain. This example is in examples/auto_memory.yml; it uses an NVIDIA NIM model, so set NVIDIA_API_KEY as well:

nat run --config_file examples/auto_memory.yml --input "I'm vegetarian and I live in Toronto."
nat run --config_file examples/auto_memory.yml --input "Suggest a restaurant for me."

Memory tools

memory:
  user_memory:
    _type: memanto

functions:
  add_memory:
    _type: add_memory
    memory: user_memory
    user_id: alex
  get_memory:
    _type: get_memory
    memory: user_memory
    user_id: alex
  delete_memory:
    _type: delete_memory
    memory: user_memory
    user_id: alex

A fixed user_id suits a single-user deployment. For multiple authenticated users, set a trusted user_id_resolver instead, as described in the toolkit's guide to authenticating memory tool users.

Tool-level user_id and user_id_resolver need NeMo Agent Toolkit 1.9 or later. On 1.8, leave them out: the agent passes user_id in each tool call.

Backend API

MemantoEditor.agent_id_for(user_id) returns the agent that holds a user's memories, so you can inspect them with the memanto CLI or UI.

Supported versions

Development

uv sync --extra test
uv run pytest

The integration tests call the live Memanto backend, including a full auto_memory_agent workflow with two users. They need MOORCHEH_API_KEY and the toolkit's --run_integration flag:

MOORCHEH_API_KEY=... uv run pytest --run_integration -m integration

Each run uses a random agent prefix and deletes the agents it creates.

Reporting bugs

• Bugs in this plugin, or in how Memanto stores and recalls memory: moorcheh-ai/memanto issues.
• Bugs in the toolkit itself, such as memory tools, the automatic memory wrapper, or plugin discovery: NVIDIA/NeMo-Agent-Toolkit issues.

License

Apache-2.0. See LICENSE.