Memory For The Agent, Not The Graph Engine
LangGraph keeps owning checkpoints, interrupts, and execution. Nowledge Mem adds the state that should survive beyond one graph or deployment: Agent context, governed Memory, cross-tool recall, and searchable Threads.
pip install nowledge-mem-langgraph
export NMEM_LANGGRAPH_APP_ID=customer-supportUse NowledgeMiddleware with LangChain create_agent, and load the bounded Mem tool set with await NowledgeClient().tools().
Identity That Survives Deployment Changes
agent_idselects the portable Mem AI Identity.host_agent_idrecords LangGraph graph/assistant provenance.space_idscopes recall and writes.- The API key authorizes access; identity fields never do.
- LangGraph
thread_ididentifies the conversation.assistant_iddoes not.
Pass invocation-specific values under context.nowledge. The connector enforces the same values on every MCP request, so model-authored tool arguments cannot switch Agent or Space.
One Thread Across Subagents
Nested subgraphs share the parent LangGraph Thread and get a nested checkpoint namespace. Mem captures the completed top-level conversation once and skips nested copies. Independently addressable subagents should use their own LangGraph thread_id and Mem Agent identity.
