Mem0 vs Letta vs Zep vs Ujex Recall: Picking an Agent Memory Layer

Pick the wrong memory layer and you'll rip it out in three months. Vectorize's writeup covers Mem0 vs Letta in detail; Letta's own essay on agent memory is canonical reading. This post adds Zep and Ujex Recall to the matrix and gives the pick-by-use-case rule.

The four shapes

Mem0

A memory service. SDK in Python and JS. Vectors via your choice of store. Hosted or self-hosted. The shape: agent calls memory.add and memory.search as plain functions; Mem0 handles chunking, embeddings, retrieval. Opinions about deduplication and update vs append.

Letta (formerly MemGPT)

An agent runtime with memory baked in. Three tiers: Core (in-context, RAM), Recall (searchable conversation, disk cache), Archival (long-term, cold storage). Letta is the agent loop, not just the memory. Self-hostable; managed Letta Code product available.

Zep

A knowledge-graph memory. Extracts entities and relations from conversation. Retrieval is graph-walk + vector. Hosted; SDKs in major languages.

Ujex Recall

Markdown files in your Cloud Storage bucket are the source of truth. Firestore holds derived metadata + vector index (Vertex AI embeddings). BM25 search baseline; vector retrieval optional. The agent reads/writes name.md files with frontmatter. Apache-2.0 SDKs.

Decision matrix

Pick Mem0 if

• You want a drop-in memory function in your existing agent loop
• Vector retrieval is the right shape for your data
• You don't need a framework (already have one or wrote your own)

Pick Letta if

• You want the framework + memory together
• You like the MemGPT three-tier model (Core/Recall/Archival)
• You're starting fresh and Letta's loop fits your problem

Pick Zep if

• Your data is naturally graph-shaped (relationships matter as much as content)
• You want managed retrieval with smart entity extraction
• Hosted-only is acceptable

Pick Ujex Recall if

• You want the agent's memory to live as .md files you can grep
• You want the bucket to be the source of truth (not a vendor's database)
• You want memory + email + audit + ingress + budgets in one project
• You want Apache-2.0 SDKs

The migration question

All four have export paths to plain JSON. Migrating between them is "export, transform, re-ingest." Vector embeddings will need recomputation (different embedding models). Plan for an afternoon, not a week.

What about Hindsight, SuperMemory, MemClaw?

Vectorize's Letta-alternatives roundup covers Hindsight, Mem0, Zep, SuperMemory. Hindsight in particular is the closest philosophical sibling to Ujex Recall — pure memory service, no opinion about your loop. MemClaw is newer; haven't tested in depth.