LangChain + NocturnusAI
The problem: ConversationBufferMemory replays every turn — cost explodes after 10 turns.
Measured on our 15-turn product-support benchmark against Claude Opus 4.
The compression happens at the Nocturnus layer — it's framework-agnostic.
Run bench.py
against your own workload.
Copy-paste install
pip install nocturnusai langchain langchain-openai
docker run -d -p 9300:9300 ghcr.io/auctalis/nocturnusai:latest
2-minute demo
15-turn support ticket with live token counts per turn.
What changes in your code
Swap ConversationBufferMemory for a call to process_turns(). Your LangChain agent now sees a compact briefingDelta instead of the full chat history.
from nocturnusai import SyncNocturnusAIClient
from nocturnusai.langchain import get_nocturnusai_tools
with SyncNocturnusAIClient("http://localhost:9300") as client:
# Compress noisy turns (user msg + tool output) into structured facts
ctx = client.process_turns(
turns=[
"User: Dashboard login failures since 8am PST, 40 users affected.",
"Tool crm_lookup: acme-corp, enterprise tier, $2.4M ARR.",
"Tool auth_audit: 287 SAML_ASSERTION_INVALID since 07:52 UTC.",
],
scope="ticket-7734",
session_id="ticket-7734",
)
print(ctx.briefing_delta)
# → "Acme Corp (enterprise, $2.4M ARR): 40-user login outage since
# 07:52 UTC. Root signal: 287 SAML_ASSERTION_INVALID errors."
# Give the agent Nocturnus query tools
tools = get_nocturnusai_tools(client)
# ... wire into AgentExecutor / create_tool_calling_agent as usualWhy it works
NocturnusAI extracts structured predicates (customer_tier(acme, enterprise), auth_error(acme, saml_invalid, 287), …) and returns only the facts reachable from your agent’s current goal via backward-chaining inference. Similarity search returns noise. This returns only what the agent needs.
What’s in the repo
support_agent.py— 3-batch support-ticket compression demo + full LangChain agent wiringbench.py— measure your own before/after tokensrequirements.txt— minimal deps
Run python bench.py after setting OPENAI_API_KEY to see per-turn usage.prompt_tokens for your workload.