Supermemory vs Mem0: An Honest Comparison
Mem0 publishes a comparison page about us. Some of it is cherry-picked, some of it is flat wrong. Here is the same comparison with the full numbers, the full context, and receipts for every claim - theirs included.
Mem0 maintains a comparison page about Supermemory on their domain (mem0.ai/compare/mem0-vs-supermemory). It ranks well, it is well designed, and parts of it are wrong. This is the same comparison, done with full numbers and receipts.
This post was reviewed against Mem0's public documentation, pricing, and their comparison page on September 17, 2026. It is published by Supermemory, so read it with that bias in mind - which is exactly why every claim below is something you can check yourself. Where we cite their page or docs, we name the source in plain text rather than linking it.
At a glance
| Supermemory | Mem0 | |
|---|---|---|
| Open source | MIT, full engine | Apache 2.0, core library |
| Self-host | Single binary, runs offline | Docker + your database |
| Free tier | $0/mo, usage credits included | 10K memories |
| Paid plans from | $19/mo usage-based | $19/mo; graph features at $249/mo Pro |
| Memory scopes | Any container tag: user, project, agent, session, org | Session, user, agent, org |
| Standing user profiles | Yes, automatic | No |
| Memory lifecycle | Versioned updates, contradiction handling, automatic forgetting | Manual lifecycle |
| Multimodal ingestion | PDFs, images, audio, URLs, files | Text and messages |
| Managed connectors | Drive, Notion, OneDrive, Gmail, GitHub, S3, web | Build your own |
| LongMemEval_s | 95% overall (Recall@15 with aggregation) | 94.4% (self-reported) |
| Retrieval latency | Sub-300ms p50, published | Not published |
The benchmark table, with the missing half
Mem0's page shows this: LongMemEval - Mem0 94.4, Supermemory 85.2. LoCoMo - Mem0 92.5, Supermemory "". BEAM - Mem0 64.1, Supermemory "".
Three corrections.
The 85.2 is not our LongMemEval result. It is one row of it. Our research page (supermemory.ai/research/longmembench) publishes the full matrix: per-category scores across single-session-user, single-session-assistant, temporal reasoning, knowledge update, abstention, and multi-session, for multiple models, with and without aggregation. The headline result: 95% overall Recall@15 with aggregation, adding roughly 720 mean tokens - a 99.4% context reduction against full-context prompting. The 85.2 Mem0 quotes is the non-aggregated gemini-3-pro row. Quoting the weakest row of a published table next to your own best number is a choice.
LoCoMo is not "~". Supermemory publishes LoCoMo results and ranks first on it. The tilde implies we never ran it. We did, and the results are public.
BEAM is Mem0's own benchmark. Not entering a vendor's home-field evaluation is not a missing result. Note also what independent leaderboards show: Mem0's 94.4 appears on no third-party LongMemEval board we could find. Ours - including the rows they quoted - are reproducible from our open benchmark harness, MemoryBench (github.com/supermemoryai/memorybench).
What Mem0 genuinely does well
Taste requires this section, and it is easy to write.
Mem0 has the largest community in agent memory - roughly 48K GitHub stars to our ~30K - and a genuinely broad set of framework integrations. The core library is Apache 2.0, simple to embed, and easy to reason about: memories as explicit objects you can inspect and manage one by one. They are YC-backed with a $24M Series A, and their managed platform carries SOC 2 and HIPAA compliance. If you want a library, not a platform, it is a reasonable one.
Where the products actually differ
Profiles vs. objects. Mem0 manages memories as discrete objects across session, user, agent, and org scopes. Supermemory does that too - container tags scope memory any way you like - and additionally maintains a standing profile per user: static facts plus dynamic, always-current context, attached to retrieval automatically. Their page lists our scopes as "Project / Container-oriented," which reads as if user-scoped memory is missing. It is not; profiles are the part they do not have.
Lifecycle vs. storage. Facts change. Supermemory versions them when they do, resolves contradictions, and forgets automatically. In Mem0, lifecycle is manual: what goes stale is yours to prune.
Ingestion surface. Mem0 ingests text and messages. Supermemory extracts from PDFs, images, audio, URLs, and files, and ships managed connectors for Google Drive, Notion, OneDrive, Gmail, GitHub, S3, and web crawling. Mem0's page flips this into our weakness - "choose Supermemory if your app depends heavily on external data ingestion" - which is an odd way to describe a superset.
Token economics. Their page lists our tokens-per-query as "Not disclosed." It is disclosed: ~720 mean tokens at Recall@15 with aggregation, on the research page, next to the scores they quoted. Retrieval latency is published too: sub-300ms p50 on the rate card.
Self-hosting. Both can be self-hosted. Ours is a single MIT-licensed binary that runs fully offline - embedded graph engine, built-in local embeddings, Ollama or any OpenAI-compatible model - with the same API after a one-line baseURL change. No enterprise agreement required, despite what some third-party listicles claim.
What the pricing comparison misses
Their page anchors "starting from $19/month" for both. The shapes differ. Mem0 meters by memory count: free to 10K memories, $19/mo to 50K, and $249/mo Pro where the knowledge graph features live. Supermemory is usage-based with included monthly credits at every tier - Free includes about $5 of usage, Pro $19/mo includes about $20, and storage and users are unlimited on every paid plan. SOC 2 and HIPAA BAAs sit on our Scale plan at $399/mo, which also includes the self-hosted option.
Which is cheaper depends on your shape: many small memories favor memory-count pricing; rich multimodal content and heavy retrieval favor usage pricing with a rate card you can model in advance.
Choose Mem0 if
- You want a small Apache 2.0 library to embed and operate yourself, and operating the backing stores is fine.
- Community size and ready-made framework integrations outweigh managed ingestion and lifecycle for you.
- Your memory needs are modest enough that memory-count pricing stays on the low tiers.
Choose Supermemory if
- You want the full engine - ingestion for any content type, extraction, profiles, lifecycle, connectors, retrieval - behind one API, with TypeScript and Python SDKs, remote MCP, and plugins for the major coding agents.
- You want standing per-user profiles, not just per-object scopes.
- You want memory that versions, resolves contradictions, and forgets on its own.
- You want the published benchmark result: 95% overall on LongMemEval_s with aggregation, first on LoCoMo, reproducible from an open harness.
- You want to self-host a single binary offline, or run managed with SOC 2 and a HIPAA BAA.
A note on corrections
We will keep this page current. If the Mem0 team spots anything here that their docs have outgrown, tell us and we will fix it. We have asked the same of them: the asymmetric benchmark table and the "Not disclosed" rows are not matters of opinion, and neither is LoCoMo.