Mem0

Mem0 pulls atomic facts out of a conversation (via an LLM call), embeds them, and retrieves the relevant ones on a later query using a mix of semantic similarity, keyword (BM25) matching, and entity linking. It's explicit fact storage with semantic search, not a system that infers behavioral patterns from usage. An Ask HN thread from a developer evaluating memory systems (2026-08) made this distinction directly: Mem0-style key-value memory captures things a user says, but doesn't learn things like 'this customer always tightens the filter threshold' from repeated behavior — that's a different problem and a different architecture. If what you need is the latter, Mem0 is the wrong shape of tool regardless of how well it does the former.

The SDK (mem0ai on PyPI/npm) is Apache-2.0 licensed and can run self-hosted as a library or Docker stack against your own vector store, with your data staying on your infrastructure. The hosted platform adds a managed API, graph memory (entity linking), a 'Dream' memory-consolidation feature, and — per the vendor's own README — proprietary retrieval optimizations not present in the open-source SDK. The repo is active: 63,617 stars, 7,443 forks, and a push as recent as 2026-08-18; the latest PyPI release (2.0.18) shipped 2026-08-11. Apache-2.0 is genuinely permissive — no field-of-use or later-relicensing clause to watch for.

Mem0 reports 92.5 on LoCoMo and 94.4 on LongMemEval, at roughly 6,900 tokens per retrieval versus 25,000+ for full-context baselines, and claims 26% higher accuracy than OpenAI's memory feature with 91% lower latency in its own published comparison. These are Mem0's own study (mem0.ai/research), not independently audited — but two things separate this from typical vendor-benchmark noise. First, the evaluation framework itself is open-sourced (github.com/mem0ai/memory-benchmarks) so the numbers are reproducible by a third party, not just asserted. Second, the vendor states plainly in its own README that these scores 'reflect Mem0's managed platform, which includes proprietary optimizations not available in the open-source SDK; open-source users should expect directionally similar gains but not identical numbers.' That's an unusually direct disclosure for a marketing surface, and it means a self-hosted deployment should not be assumed to match the published figures.

Hobby (free): 10,000 add requests/month, 1,000 retrieval requests/month, 1 project. Starter: $19/month for 50,000 add / 5,000 retrieval requests. Pro: $249/month for 500,000 add / 50,000 retrieval requests, unlimited projects, graph memory, and Dream. Enterprise is custom-priced with unlimited requests, on-premises deployment, audit logs, and SSO. All tiers advertise 'unlimited end users' — the actual constraint is request volume, not seat or user count. The jump from Starter to Pro (13x the price for 10x the retrieval quota) is the tier a growing team will hit fastest; below that the free tier is generous enough for real prototyping, not just a toy quota.

Mem0's trust center (trust.mem0.ai, verified today) lists HIPAA readiness, GDPR, and a Vanta-issued SOC 2 Type I badge. Type I attests controls are designed correctly at a single point in time; it is a materially lighter bar than Type II, which attests the controls operated effectively over a period (commonly 6-12 months). The vendor's own page does not currently show a Type II badge, so treat SOC 2 status as 'in progress toward the stronger attestation,' not as equivalent to it.

Mem0 is a Y Combinator S24 company that raised $24M in combined seed and Series A funding, announced 2025-10-28. The Series A was led by Basis Set Ventures with participation from Peak XV Partners and GitHub Fund, alongside strategic angels including Datadog's and Supabase's CEOs. This is not a side project — it's a funded startup with runway, which matters for anyone evaluating whether to build on its hosted platform long-term.

The clearest third-party corroboration found is an AWS Machine Learning blog post describing Trend Micro combining Amazon Bedrock, Neptune, OpenSearch, and Mem0 for an internal chatbot, and a Groq customer-story page describing a Mem0 integration on GroqCloud — both independently published, not Mem0's own case studies. Mem0's own blog separately publishes case studies for RevisionDojo and OpenNote (both claiming 40% token-cost reduction), which should be read as vendor-supplied rather than corroborated. No review base was found on G2 or Trustpilot with enough volume to characterize — this is a young product with limited independent user-review evidence yet, which is normal at this stage rather than a red flag.