OpenAI Agents SDK Integration

Engram can be used from OpenAI Agents SDK workflows through function tools backed by the Engram HTTP API. This does not require MCP.

Start a local Engram HTTP server:

engram serve --http

Install the OpenAI Agents SDK in the application that will run the agent:

pip install openai-agents

Usage

from agents import Agent, Runner

from engram.integrations.openai_agents import create_engram_tools


tools = create_engram_tools(
    base_url="http://127.0.0.1:7474",
    api_key="ek_live_YOUR_INVITE_KEY",
    default_scope="auth",
)

agent = Agent(
    name="Engram-aware agent",
    instructions=(
        "Query Engram before making architecture decisions. "
        "Commit only verified discoveries, decisions, and corrections."
    ),
    tools=tools,
)

Runner.run_sync(agent, "What do we know about auth?")

The integration exposes three tools:

Tool	Purpose
engram_query	Read verified workspace facts relevant to a topic
engram_commit	Commit a verified fact to Engram
engram_conflicts	Review open or resolved conflicts before important decisions

Memory Model

The OpenAI Agents SDK Session memory stores conversation turns. Engram stores verified shared workspace facts. Keep those responsibilities separate:

• use SDK sessions for short-term conversation history
• use Engram for durable team knowledge that should survive across agents and sessions
• do not commit raw chat transcripts to Engram

Without an Invite Key

For local development without auth:

tools = create_engram_tools(base_url="http://127.0.0.1:7474")

For team mode, pass the invite key as api_key; it is sent as a Bearer token to Engram's REST API.