Author(s): Chew Loong Nian – AI ENGINEER

Originally published on Towards AI.

Mem0 vs Zep vs Letta: A Folder of Text Files Shouldn't Beat the 61K-Star Memory Layer

The most-installed agent memory layer on GitHub has passed 61K stars. Letta beat its published LoCoMo score by dumping the conversation into a plain file and handing the agent grep: 74.0% against Mem0's 68.5%. No knowledge graph, no vector store tuned for the task, no memory product at all.

The article argues that the agent-memory “leaderboards” vendors rely on are not independently reproducible, contain methodological and even arithmetic errors (including a reported LoCoMo inflation of 25.56 percentage points), and measure the wrong thing—especially because LoCoMo mostly tests short-range retrieval under benign conditions and largely ignores knowledge updates. The author walks through the chain of disputes among Mem0, Zep, and rebuttals, explains why filesystem-based retrieval can outperform semantic ADD-only memory on freshness when recency is handled correctly, and introduces a simple 20-probe staleness test you can run without APIs or LLMs to measure whether a system returns corrected facts or repeats stale ones. The piece concludes with practical selection guidance (Mem0 for cross-session personalization today, Zep/Graphiti for audit-like temporal knowledge, Letta for long-running autonomous agents, or a plain folder/files approach for many coding-terminal use cases) and a verdict to stop buying based on public scores and instead run domain-specific contradictions against your own workload.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI