Renor vs Mem0 vs Zep: memory for coding agents

Developers who run Claude Code, Codex and Cursor often look at the same names when they want durable memory: Mem0, Zep and Renor. We build Renor, so read this with that in mind. This page compares what each one publishes, and says where a source does not say enough to compare.

Answer models, judges and retrieval settings differ from page to page, so the scores below are not directly comparable. Competitor figures are quoted as the provider published them, read on October 28, 2026, and none was re-run by Renor.

What Renor is

Renor is one memory for all your coding agents, shared over MCP at https://mcp.renor.cloud/mcp. From how it works and the quickstart:

• Agents save with remember or save_memory and recall with get_context.
• Memories are plain Markdown, and the whole workspace exports as Markdown.
• A new fact with the same subject, predicate and scope supersedes the old value; the old one stays in history.
• The engine is open source under Apache-2.0 (renorai/renor-engine) and can be self-hosted with npx -y renor-engine mcp.
• It does not record chats automatically, does not merge contradictory plain notes, and has no shared workspace for several people yet.
• Hosted pricing: Free with 3,000 memory calls a month, Pro at US$10 a month (pricing).

For Mem0 and Zep, use their own pages for features and current pricing: mem0.ai and getzep.com. We do not summarise them here, because those pages change.

LongMemEval-S (500 questions)

Renor's deep search uses --top-k 50 and a 4,000-token ranked-expansion budget. That setting is off by default: an install with default settings and the same reader scores 94.4%. The settings were developed on these same 500 questions, and each Renor figure is one run.

LoCoMo (categories 1 to 4, 1,540 questions)

Renor's two rows differ in reader, judge and page size at once, so the gap between them is not a retrieval gain by itself. And neither Renor row shares both reader and judge with Zep's, so no row here ranks Renor above or below Zep or Mem0.

BEAM

BEAM tests memory at 100K, 500K, 1M and 10M tokens. Scores are nugget scores from 0 to 100, and each vendor grades with its own judge.

Other vendors publish BEAM figures, for example Hindsight (April 2, 2026), which reports a higher score than Renor's 60.2 at 10M tokens. We do not copy competitor BEAM numbers into this table: each vendor grades with its own judge, and we have not recorded a source that names the answer model and judge. Read the vendor's own page, and expect Renor to be behind at 10M.

Tokens read per question

On LongMemEval-S, the text sent to the answering model averages 5,537 tokens for Renor's deep-search page, against 6,787 for Mem0 as printed on Mem0's evaluation page at its top-200 budget. A whole conversation is about 115,000 tokens. This row compares the same quantity; the accuracy rows do not.

Choosing for coding agents

Choose Renor if you want:

• One MCP memory shared by Claude Code, Codex and Cursor today.
• Markdown export and a fact ledger that keeps history when a value changes.
• A free hosted tier or an Apache-2.0 self-host.

Look at Mem0 or Zep if you need their platform features, SDKs or products beyond a coding-agent MCP memory, and check their live docs and pricing first.

Whichever you pick, prefer a page that names the answer model, the judge and the retrieval settings. A headline percentage alone is not enough. We explain how we measure in How we measure Renor.

Try Renor

Create a free account, follow the quickstart, or read the research page.