Every team building AI agents eventually wants to feed their accumulated chat history into a knowledge graph. Most do it wrong the first time: they flatten conversation trees into linear transcripts and lose the branch structure where the real decisions lived. The graph then confidently links assistant claims from sibling branches as if they happened in one conversation, and agents start retrieving contradictory advice as if it were consensus.
Dan Mercede built and runs the ChatGPT-to-Cognee ingest pipeline that feeds a production Neo4j 5 + Cognee MCP knowledge layer for an estate of AI coding agents. In this 25-minute talk, he will walk attendees through the implementation that makes branched conversation ingest work in production. The core moves: full traversal of ChatGPT's mapping dictionary instead of current-node following, per-branch serialization that preserves decision lineage, content-addressed identity for branch paths, and idempotent dedupe across re-runs, plus the integration boundary with Cognee's cognify pipeline for downstream entity and relationship extraction.
He'll show the Neo4j data model the pipeline produces, the Cypher queries that surface branch-aware provenance, the dedupe pattern that prevents conversation-export and historical-archive sources from double-counting, and the lessons from a live Neo4j storage-identity migration that preserved every original node and relationship identity.
You will leave with a working blueprint for ingesting branched LLM conversation history into a knowledge graph, the patterns to keep it idempotent and re-runnable, and the failure modes to design against before you hit them in production.
Speaker
Dan Mercede
AI Systems Architect, Agent Runtimes and Retrieval Pipelines
Dan Mercede is an AI systems architect who builds the infrastructure agents run on: ingest and retrieval pipelines, agent runtimes, and the layer that makes agent behavior auditable.
He built and runs the conversation-ingest pipeline behind a production Neo4j knowledge layer serving an estate of AI coding agents, and is the author of mcp-context-budget (github.com/OrionArchitekton/mcp-context-budget), an open-source tool for measuring and enforcing MCP tool-surface budgets. His work centers on one question: can you prove what your agent actually did? In practice that means append-only receipts, fail-closed execution, and pipelines built to be idempotent and re-runnable. More of the work, with evals published including the misses, at danmercede.com/works.
Before AI infrastructure, he spent four years as a general manager with full P&L ownership, an operator's lens he still builds with.
