NVIDIA NeMo Agent Toolkit + MemorySync

Overview

Install nat-memorysync and `_type: memorysync_memory` becomes available in every NeMo Agent Toolkit workflow YAML — a real entry-point plugin, not example code to vendor (the toolkit’s Mem0 and Zep editors live in its repo as examples). It works with both wiring modes: NVIDIA’s built-in add_memory/get_memory tools, and the automatic auto_memory_agent wrapper that stores every turn and enriches every prompt with no tools and no prompt changes. The editor is engineered against the in-repo editors’ sharp edges: search without user_id raises a ValueError naming the kwarg (the Mem0 editor throws a bare KeyError), your metadata dict is never mutated (the Mem0 editor pops keys out of it), search returns one MemoryItem per fact with similarity_score populated (the Zep editor returns a single joined blob; the Mem0 editor discards scores), and remove_items with no kwargs raises instead of silently doing nothing. Deletes are session-scoped by default — wiping a user’s entire memory requires the explicit scope="user" opt-in. Rows are always keyed by each item’s user_id, so an unset conversation id can never mix users (the in-repo Zep editor routes everyone to one shared "default_zep_thread"). Recall runs under a hard 1.2s budget and fails open, so automatic memory can never stall a turn; deterministic idempotency seeds make RetryMixin retries converge on one stored row.

Setup status and requirements

Supported version

nat-memorysync 1.0.0 (PyPI)

Last setup review

2026-08-24

Permissions

A MemorySync API key with read and write scopes.

Limits

Requires nvidia-nat-core 1.5+ (Python 3.11+, the toolkit’s own floor).

Capabilities

• Real pip-installable plugin via the nat.components entry point — not vendored example code
• Both wiring modes: add_memory/get_memory tools and the automatic auto_memory_agent
• One MemoryItem per fact with similarity_score — no joined blobs, no discarded scores
• Loud typed contracts — ValueError naming the missing kwarg, never a bare KeyError
• Caller items and metadata never mutated (copy-first, tested)
• Session-scoped deletes by default; whole-user wipe is an explicit scope="user" opt-in
• Bleed-safe by construction — rows always keyed by item user_id
• Hard 1.2s recall budget, fail-open both directions; RetryMixin-safe idempotent writes

Quick Start

memory:
  saas_memory:
    _type: memorysync_memory   # key from MEMORYSYNC_API_KEY

functions:
  add_memory:
    _type: add_memory
    memory: saas_memory
    description: Save any user preference or fact.
  get_memory:
    _type: get_memory
    memory: saas_memory
    description: Recall saved user preferences and facts.

workflow:
  _type: react_agent
  tool_names: [add_memory, get_memory]
  llm_name: my_llm

Use Cases

Explore More