neruva-langchain 0.2.1

Drop-in LangChain integration for Neruva agent memory substrate. v0.2.0 adds AUTO-PILOT: NeruvaAutoPilot (classify intent + reflect + extract via your LangChain LLM) and NeruvaAutoPilotCallback (BaseCallbackHandler). Plus NeruvaChatMessageHistory + NeruvaContextRetriever for memory + KG recall, NeruvaCodeGraph (5 code nav methods), NeruvaCognitive (13 cognitive primitives). Pattern-C: substrate stays $0/call.

neruva-langchain

Drop-in LangChain integration for Neruva agent memory + reasoning substrate. Three wrappers cover the common LangChain plug points: chat history, document retriever, and (new in 0.2.0) auto-pilot intent routing + reflection.

pip install neruva-langchain

What's new in 0.2.0 — Auto-pilot

Two new exports complete the auto-pilot integration. The agent now proactively uses the right Neruva cognitive tool (counterfactual / analogy / theory-of-mind / rule induction / EFE planning / causal / etc.) based on the user's intent — without you wiring each one.

• NeruvaAutoPilot — framework-agnostic core. Pass your LangChain LLM; get back classify_intent(), reflect(), and extract_facts() methods that round-trip through the substrate's canonical prompts (Layer 2 / 3 / 1 respectively).
• NeruvaAutoPilotCallback — BaseCallbackHandler subclass that auto-fires intent classification on on_chain_start and (opt-in) reflection on on_chain_end after every 10 turns. Drop into any chain via callbacks=[NeruvaAutoPilotCallback(...)].

Plus NeruvaCodeGraph (5 sub-ms code-navigation methods — callers / callees / class_of / module_of / imports) and NeruvaCognitive (13 cognitive-primitive methods — counterfactual rollout / theory-of-mind / schema lifting / EFE planning / continual K-gram / hierarchical chunking / rule induction). Thin wrappers over NeruvaClient so they're callable from anywhere in your chain.

Pattern-C: substrate stays $0/call. The classification + extraction work runs in YOUR LangChain LLM's turn.

from langchain_anthropic import ChatAnthropic
from neruva_langchain import NeruvaAutoPilot, NeruvaAutoPilotCallback

llm = ChatAnthropic(model="claude-3-7-sonnet-20250219")
ap = NeruvaAutoPilot(api_key="nv_...", namespace="my_app", llm=llm)

# Layer 2 — classify a user message
result = ap.classify_intent("what if we had picked Mailgun?")
# {'primary_intent': 'counterfactual', 'confidence': 0.92,
#  'suggested_tool': 'agent_counterfactual_rollout', ...}

# Layer 3 — reflect over recent turns + auto-write durable records
ap.reflect(recent_turns=[
    {"role": "user", "text": "ship the ticket system"},
    {"role": "assistant", "text": "shipped 5 phases including notes + audit"},
])
# Auto-writes decisions/facts/mistakes to substrate.

# Or auto-fire via callback in any LCEL chain / agent:
chain = prompt | llm | parser
result = chain.invoke({"q": "..."}, config={"callbacks": [
    NeruvaAutoPilotCallback(api_key="nv_...", namespace="my_app", llm=llm,
                            enable_route=True, enable_reflect=True),
]})

What's new in the substrate (v0.5.7, May 2026)

The substrate this adapter wraps has gained a lot since the last release. All of it is available via the same agent_* API the wrappers already call, so existing code keeps working — new capabilities just become available.

• Deterministic replay — every query is bit-identical across reruns from the same seed. Replay any past state for audit or debugging.
• Typed-shape context — pull structured JSON from records with per-field citations. {question, shape: {field: type}} → typed result without an LLM at query time. Natural fit for tool-calling chains that need a specific output schema.
• Tenant-specific PII rules — register your custom ID formats (employee codes, patient codes, order IDs) from 3-5 examples. The substrate redacts them automatically. Sub-microsecond per span.
• Depth-unlimited nested-belief tracking — store and retrieve chains like Alice → Bob → Carol thinks X at any depth, with inner-position-swap rejection.
• Counterfactual rollouts — "what if action k had been a' instead?" Replay an action sequence with one step substituted.
• Active inference planning — score candidate action sequences by KL distance to a goal-marginal. Caller-owned dynamics.
• Continual K-gram learning — provable no-forgetting via integer-add commutativity; repeated train() calls accumulate.

NeruvaChatMessageHistory

Auto-records every turn into the Neruva Records substrate. Drop into any LangChain primitive that accepts a BaseChatMessageHistory:

from neruva_langchain import NeruvaChatMessageHistory

history = NeruvaChatMessageHistory(
    api_key="nv_...",          # or env NERUVA_API_KEY
    namespace="user_alice",    # one per user / session
)

history.add_user_message("My name is Alice and I live in Toronto.")
history.add_ai_message("Nice to meet you, Alice!")

# Later — even after process restart, substrate-backed:
print(history.messages)

Use with RunnableWithMessageHistory (modern pattern)

from langchain_anthropic import ChatAnthropic
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain_core.runnables.history import RunnableWithMessageHistory
from neruva_langchain import NeruvaChatMessageHistory

prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful assistant."),
    MessagesPlaceholder(variable_name="history"),
    ("human", "{input}"),
])
chain = prompt | ChatAnthropic(model="claude-opus-4-7")

chain_with_history = RunnableWithMessageHistory(
    chain,
    lambda session_id: NeruvaChatMessageHistory(namespace=session_id),
    input_messages_key="input",
    history_messages_key="history",
)
chain_with_history.invoke(
    {"input": "What did I tell you about my project last week?"},
    config={"configurable": {"session_id": "user_alice"}},
)

NeruvaContextRetriever

BaseRetriever for RetrievalQA chains. Returns Document objects sourced from federated agent_recall:

from neruva_langchain import NeruvaContextRetriever
from langchain.chains import RetrievalQA
from langchain_anthropic import ChatAnthropic

retriever = NeruvaContextRetriever(
    api_key="nv_...",
    namespaces=["session_a", "session_b"],   # multi-session fan-out
)
qa = RetrievalQA.from_chain_type(
    llm=ChatAnthropic(model="claude-opus-4-7"),
    retriever=retriever,
)
qa.invoke("Where does Alice work?")

Why use Neruva instead of LangChain's built-in memory?

Feature	LangChain default	Neruva
Persists across process restart	Manual setup	Built-in (GCS-backed)
Cross-session recall	No	Yes via namespaces=[...]
Fact extraction (KG)	No	Auto (hd_kg_extraction_prompt + caller LLM)
GDPR forget by user	Manual	user_id= auto-folds, one-call forget
Determinism / replayability	No	Bit-identical from seed
Portability	Pickle	.neruva zip container

Get an API key · Docs · Status