iCog vs Mem0 vs Letta vs Zep: choosing an AI memory layer
Short version: Mem0, Letta, and Zep are developer infrastructure — SDKs you wire into agents you build. iCog is a personal, MCP-native memory layer that plugs into the AI tools you already use (Claude, Cursor, ChatGPT, Codex) and adds a reflection/cognition layer on top. If you're building an agent product, look at the first three. If you want one memory across your own AI tools, that's iCog's lane.
The "AI memory" space split into two layers in 2026, and it's worth knowing which one you're shopping in before you compare features.
The two layers
Developer infrastructure — Mem0, Letta, Zep, Cognee, Supermemory. You install an SDK, call an API, and your agent gets memory. These compete on benchmarks (LongMemEval, LoCoMo), graph sophistication, latency, and enterprise compliance. They sell to teams building agents.
Personal cross-agent memory — products that give you one memory across every AI you touch. Newer, less consolidated, and where iCog lives. The pitch isn't "add memory to your agent," it's "your AI finally knows you."
The rundown
Mem0
The category default. A universal memory layer for agents, broad framework integrations, strong research output, real funding and adoption. If you're building an agent and want the most-integrated, best-documented building block, Mem0 is the safe default. It's infrastructure, though — a store you wire in, not a product that knows you.
Letta (formerly MemGPT)
Originated much of this field with the MemGPT paper. Letta's strength is transparent, editable memory blocks and a stateful-agent runtime — and it's increasingly a full agent platform, not just memory. Great if you want an opinionated runtime for stateful agents. Heavier if all you need is memory.
Zep
The enterprise pick. A temporal knowledge graph (Graphiti), sub-200ms recall, and the compliance checklist (SOC 2, HIPAA, GDPR) that wins B2B deals. Built for customer-support and sales AI at scale. Powerful and serious — and aimed squarely at enterprises, not individuals.
iCog
MCP-native and personal. Instead of an SDK you embed, iCog is a memory server you connect to the tools you already use — so Claude, Cursor, ChatGPT, and Codex share one long-term memory of you. On top of storage it runs a cognition layer: it reflects on conversations, consolidates memories (we call it dreaming), and forms a model of who you are by memory type — episodic, semantic, procedural, foundational. The differentiator isn't a benchmark score; it's that the memory thinks about you, not just stores you.
Which should you pick?
| If you're… | Look at |
|---|---|
| Building an agent product for a dev team | Mem0 (broad, integrated) |
| Building stateful agents and want an opinionated runtime | Letta |
| Shipping enterprise AI that needs compliance + scale | Zep |
| A builder or power-user who wants your own AI tools to remember you | iCog |
The honest take
Don't pick a memory layer on a benchmark leaderboard if you're not running that benchmark's workload. The infra players are in a LongMemEval/LoCoMo arms race that mostly matters if you're shipping agents at scale. For a person who lives across several AI tools and is tired of re-explaining themselves to each one, the question isn't "which scores highest on multi-hop recall" — it's "which one gives me one memory everywhere, and actually understands me over time."
That's the bet iCog makes. Not smart. Just knows you.
One memory across Claude, Cursor, ChatGPT, and Codex. Try iCog →