Skip to content

LANGCHAIN_GUIDE.md

Latest commit

 

History

History
106 lines (77 loc) · 2.82 KB

LANGCHAIN_GUIDE.md

File metadata and controls

106 lines (77 loc) · 2.82 KB
Raw

LangChain 1.x Integration

SCM integrates through LangChain's middleware hooks. It runs in-process with the agent and requires no SCM server or SCM endpoint.

Fixed User

from langchain.agents import create_agent
from scm import SCM

memory = SCM(user_id="alice")
agent = create_agent(
    model=model,
    tools=[...],
    middleware=[memory.langchain(max_memories=8, max_context_chars=4000)],
)

result = agent.invoke({
    "messages": [{"role": "user", "content": "Keep my answers concise."}]
})

One-Call Helper

from scm.langchain import create_scm_agent

agent = create_scm_agent(
    model=model,
    tools=[...],
    user_id="alice",
    system_prompt="You are a technical support agent.",
)

The returned agent exposes its runtime as agent.scm for lifecycle and administration calls.

Multi-User Runtime Context

from scm import SCM, SCMContext
from scm.langchain import create_scm_agent

memory = SCM(namespace="support")
agent = create_scm_agent(model=model, tools=[...], memory=memory)

config = {"configurable": {"thread_id": "langgraph-thread-19"}}
context = SCMContext(user_id="alice", thread_id="ticket-19")

agent.invoke(
    {"messages": [{"role": "user", "content": "My account uses SSO."}]},
    config=config,
    context=context,
)

Use the same SCMContext with invoke, ainvoke, and streaming methods. A LangGraph checkpointer's thread identity and SCM's application user identity serve different purposes; SCM records both without letting the model choose either user.

Middleware Semantics

Before the model call, SCM searches the current application user's memory and adds a bounded <scm_memory> block to the system message. The block is labeled as untrusted user data, so stored prompt-injection text is context rather than instructions.

Human messages are persisted before model execution. Writes fail closed: a durability error aborts the model call with SCMWriteError. Recall fails open: the model can continue and scm_diagnostics records the recall failure class.

Retries and graph resumptions use a hash of user, thread, and message identity, so replaying the same message does not create duplicate memories.

Canonical Tools

Middleware contributes exactly five tools:

  • add_memory
  • search_memory
  • sleep
  • wake_summary
  • forget

Their schemas never contain user_id. The tools resolve identity from the LangChain runtime context.

Model Reuse

SCM attempts to reuse the attached LangChain model for structured concept extraction. If that model does not support structured output or extraction fails, SCM falls back to offline heuristics. A second provider key is not required.

Shutdown

Long-running applications should close SCM during graceful shutdown:

agent.scm.close()

This drains background work and commits the final lifecycle snapshot.