AI For Dev

summary-nate-herk-zep-memory

2 min read

Unlock the Next Evolution of Agents with Human-like Memory (n8n + Zep)

Author / channel: Nate Herk Format: video Source: Original Published: 2025-07-14

Summary

Nate Herk demos giving an n8n AI agent long-term memory via Zep’s relational graph — entities and relationships, not just turn history. Then identifies the cost trap: naive Zep integration sends all facts to the LLM every turn, blowing up token usage. Builds a hybrid pattern that uses HTTP requests to (a) pull the 3 most relevant facts from Zep (relevance threshold 0.7) for the current query and (b) use Postgres for short-term conversation history (because Zep’s native context API can’t reverse-order pulls). Token usage drops from ~2,400 to ~1,000 per turn for the same query.

Key Points

  • Why simple memory isn’t enough — n8n’s built-in simple memory only stores the last N turns. No long-term entity knowledge, no relationships.
  • Zep relational graph — each user gets a graph of entities (yourself, places, interests, tools) with typed relationships (“uses”, “lives in”, “likes watching”). Each chat populates the graph automatically.
  • The cost trap — n8n’s native Zep integration sends the entire user graph + full session history every turn. Demo: 2,438 tokens for a single turn against a small graph. Scales badly.
  • Method 1: HTTP-based selective fact pull — instead of native Zep node, use HTTP request to Zep’s search endpoint with query, userId, limit: 3, min_score: 0.7. Then a JS code node cleans up the response. Result: 3 most-relevant facts only.
  • Method 2: Hybrid Zep + Postgres — Zep for the user graph (long-term), Postgres for context window (short-term). Reason: Zep’s session-message endpoint pulls oldest-first with no reverse parameter. Postgres handles the recency requirement cleanly.
  • Token savings — same prompt: 2,438 tokens (naive Zep) → 1,045 tokens (selective pull). ~2.4× reduction.
  • Session ID = identity hook — using Telegram chat ID, email address, or user ID as the Zep session ID gives each end user their own persistent graph. Powerful for onboarding bots, tutoring agents, customer-specific assistants.
  • Code-from-Claude trick — Nate’s pattern when an n8n code node is needed: paste sample input JSON into Claude, describe the desired output shape, paste the returned code. He doesn’t write JS himself.

Connected Pages

  • n8n — backbone
  • zep — entity
  • Nate Herk — channel hub (now 2 sources)
  • open-brain — alternative agent-memory architecture (Postgres + MCP)
  • OpenBrain Architecture — comparison point: Zep is hosted-graph, OpenBrain is DIY-Postgres

Graph View

Backlinks