langchain-quickstart

LangChain Quickstart

A returning customer keeps their plan, preferences, and prior context across turns — delivered to a LangChain chain via a small BaseMemory subclass backed by Statewave.

What this demo shows

Capability	How it's demonstrated
Drop-in LangChain memory	StatewaveMemory(BaseMemory) plugs into any chain via the memory= kwarg
Token-bounded, ranked context	Each turn gets a fresh Statewave bundle sized to max_tokens, biased toward the current input
Durable across runs	save_context records every turn as an episode — restart the script and the model still knows you
No vector DB plumbing	Statewave's compile → retrieve loop replaces the usual "embed + Pinecone" stack inside the chain
Provenance preserved	Memory in the prompt is the same provenance-tagged bundle the rest of Statewave returns

Prerequisites

A running Statewave server at http://localhost:8100 — the docker-compose.yml in the examples root brings up Postgres 16 + pgvector and the API together:

docker compose up -d         # from statewave-examples/

Dependencies (no framework deps are pinned in the core Statewave SDK — install them at the example level). LangChain 1.x removed the BaseMemory class this adapter extends, so pin the 0.3 line:

pip install "statewave>=0.10.0" "langchain>=0.3,<1" "langchain-openai<1"
export OPENAI_API_KEY=sk-...

Run

python langchain_quickstart.py

The demo seeds Alice's plan and contact preference as episodes, compiles them into memories, and then asks a follow-up question. The chain pulls the right slice of memory via StatewaveMemory and the model answers grounded in it.

The adapter

adapter.py is the entire integration — about 40 lines:

• StatewaveMemory(BaseMemory) — the LangChain memory class.
  • load_memory_variables({"input": ...}) → fetches a Statewave bundle for the current input and exposes it as {memory_key: <context>} (default key: statewave_context).
  • save_context(inputs, outputs) → appends the (user, assistant) turn as a Statewave episode.
  • clear() is intentionally a no-op — Statewave is append-only, and accidentally dropping durable memory because a chain was reset is the wrong default. Call client.delete_subject(subject_id) explicitly when you really mean it.

Reference it in your prompt template via the {statewave_context} placeholder:

prompt = PromptTemplate(
    input_variables=["statewave_context", "input"],
    template="{statewave_context}\n\nUser: {input}\nAssistant:",
)

Smoke test

The adapter's wiring is covered by a small mock-based test — no LLM, no Statewave server, only LangChain installed:

pip install langchain
pytest test_adapter.py

See also

• Statewave Python SDK — the underlying StatewaveClient.
• Getting started — bring a server up in 5 minutes.
• ../support-agent-python/ — the same retrieve-then-respond pattern without LangChain.