VEKTOR vs Zep
AI Agent Memory (2026)

Both give your agents persistent memory across sessions. The architecture, ownership model, and pricing are completely different. Here's the honest breakdown.

Local-first · Node.js · MCP-native VEKTOR $9/mo flat · no usage fees
VS
Cloud · Python-first · Graphiti engine Zep Usage-based cloud + self-host enterprise tier

Quick comparison

Feature
VEKTOR
Zep
Storage
Local SQLite — your machine
Zep Cloud or self-host (enterprise)
Data leaves server
Never
Yes (cloud tier)
Recall latency
8ms avg · <50ms p95
100–400ms (cloud round-trip)
Pricing
$9/mo flat
Usage-based (scales with queries)
MCP server
✓ Native — Claude, Cursor, Windsurf
Not natively — requires wrapper
Primary language
Node.js / TypeScript
Python-first (TS SDK available)
Graph architecture
MAGMA — 4 layers (semantic, causal, temporal, entity)
Graphiti — temporal knowledge graph
Temporal reasoning
Temporal layer + decay scoring
Graphiti excels — validity windows per fact
Auto-curation
AUDN — ADD/UPDATE/DELETE/NO_OP per fact
Graphiti conflict resolution
Background compression
REM cycle — 50:1 ratio async
No equivalent
Embedding cost
$0 — uses your existing LLM key
Billed per embedding on cloud
Open source
Vex + Vek-Sync OSS / SDK commercial
Zep OSS on GitHub

Architecture

VEKTOR — MAGMA Graph

VEKTOR stores memory in a local SQLite database as a 4-layer associative graph. Every remember() call runs through AUDN — an autonomous curation layer that decides whether to ADD a new node, UPDATE an existing one, DELETE a contradiction, or do nothing. The graph self-organises via Zettelkasten-style edges typed as SUPPORTS, EXTENDS, CONTRASTS, or PREREQUISITE.

A background REM cycle (7 phases) compresses raw memory fragments into distilled insights while your agent is idle — 50 fragments become 3 core signals. Nothing blocks the agent.

All of this runs locally. No cloud call, no API roundtrip for memory. Recall averages 8ms.

Zep — Graphiti Engine

Zep's Graphiti engine is purpose-built for temporal reasoning. Every stored fact is a knowledge graph node with a validity window — "User prefers Python (as of March 2026)" is stored with a time bound, not just a string. When facts change, old nodes are closed out and new ones opened.

This makes Zep particularly strong for agents that need to reason about how things changed over time — a gap where most memory layers fall short, including basic vector stores.

The cloud tier handles embedding and retrieval on Zep's infrastructure. Self-hosting is available but scoped to enterprise plans.

Pricing

VEKTOR is $9/month flat — one price regardless of how many queries your agent makes, how many memories it stores, or which LLM provider you use. There are no per-query fees, no embedding bills, no egress charges. The SDK uses whatever LLM API key you already have configured.

Zep is usage-based on the cloud tier. The free tier covers low-volume experimentation, but production workloads — agents that query memory hundreds of times per session — accumulate costs that scale with usage. Enterprise self-hosting requires a contract.

For teams running multiple agents or high-volume workflows, VEKTOR's flat pricing model is significantly more predictable.

MCP support

VEKTOR ships a native MCP server. Adding it to Claude Desktop, Cursor, Windsurf, or VS Code is a one-line config change. The MCP tools — vektor_recall, vektor_store, vektor_graph, vektor_delta — are available immediately with no additional setup.

Zep does not currently ship a first-party MCP server. Integration into MCP-native environments requires a custom wrapper or HTTP bridge. For teams building on Claude Desktop or Cursor specifically, this is a meaningful friction difference.

When Zep is the better choice

When VEKTOR is the better choice

Bottom line

These are two well-built tools solving the same problem from different angles. Zep's Graphiti engine leads on temporal reasoning — if your agent needs to track how facts change over time, it's the most architecturally capable option in the market. VEKTOR leads on everything else that matters for Node.js / MCP-native / privacy-first / cost-predictable workloads: local-first, 8ms recall, flat pricing, native MCP, and the only background compression system in the category.

If you're unsure, the install is 2 minutes and the first month is $9.

Try VEKTOR

Local-first. 8ms recall. MCP-native. $9/month flat.

Get VEKTOR Read Docs