Framework runtimes are different from host agents. agentmemory connect <agent>
mutates a local agent config file, but OpenAI Agents, AutoGen, CrewAI,
LangGraph, and similar runtimes live inside user applications. The safe default
is descriptor-only setup guidance: no project files are rewritten and no vendor
client is imported by agentmemory.
Use the helper:
agentmemory connect frameworks
agentmemory connect frameworks langgraph
agentmemory connect frameworks --jsonThe descriptor registry lives in src/cli/connect/frameworks.ts.
| Name | Runtime | Primary setup surface |
|---|---|---|
openai-agents |
OpenAI Agents SDK | REST lifecycle hooks |
autogen |
Microsoft AutoGen | MCP tools or REST lifecycle hooks |
crewai |
CrewAI | MCP/custom tool plus REST capture |
langgraph |
LangGraph | REST node hooks, optional MCP tools |
pydantic-ai |
Pydantic AI | Typed dependency/tool using REST |
semantic-kernel |
Semantic Kernel | Kernel function or MCP tool |
llamaindex-workflows |
LlamaIndex Workflows | REST workflow hooks |
mastra |
Mastra | MCP/custom tool plus REST capture |
Start the daemon first:
agentmemorySet runtime env:
AGENTMEMORY_URL=http://localhost:3111
# AGENTMEMORY_SECRET=only-if-configuredUse the framework's own short-term memory, checkpointing, or workflow state for in-run behavior. Use agentmemory for durable facts and cross-session recall:
- Call
/agentmemory/contextbefore a run, graph entry, or retrieval-heavy node when the app wants recalled project context. - Call
/agentmemory/rememberor/agentmemory/observeafter durable user, tool, handoff, task, or workflow events. - Call
/agentmemory/agent-eventswith framework-native ids such as run, thread, team, task, node, tool-call, or checkpoint ids. - Use the MCP bridge only when the framework should expose memory tools directly to an agent.
These descriptors are intentionally metadata and setup helpers, not automatic installers. A future framework-specific package can consume the same descriptor names without changing the host-agent adapter registry.