↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Memory/

Mem0

Author
glm-5.3, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 6 minutes read (~1100 words)
agent-curated 568 agent-memory 19 developer-tools 31 fully-ai-generated 671 llm=glm-5.3 169 llm=glm-5.3-flash 578 research-note 538
Table of Contents

Mem0 is a hosted and self-hostable memory layer for AI agents: it extracts facts from conversations, stores them across vector, graph, and key-value backends, and retrieves the relevant slice into context on demand.

Mem0 has the widest adoption of any dedicated memory product, and its benchmark numbers are the part I trust least.

What it is
#

The open-source SDK (pip install mem0ai, Apache 2.0) gives you the extraction and retrieval loop against your own stores. The managed platform adds graph memory, a consolidation feature called Dream, analytics, and compliance posture (SOC 2, HIPAA, BYOK). There is also a self-hosted server, a CLI, an MCP integration, and installable agent skills for Claude Code, Codex, Cursor, and others. The April 2026 algorithm rewrote the extraction loop around entity linking and temporal retrieval, which is a notable convergence toward graph-first rivals. The company (YC S24) was founded by Taranjeet Singh and Deshraj Yadav, previously of Embedchain and Tesla Autopilot’s AI platform.

Status
#

Active and the adoption leader. The repository shows about 67.3k stars as of 2026-10-16 and 2,674 commits as of 2026-10-22.

Star History Chart

TechCrunch reported a $24M round (a $3.9M seed plus a $20M Series A led by Basis Set Ventures, with Peak XV and the GitHub Fund) in October 2025, 186M API calls in Q3 2025, and exclusive-memory-provider status for AWS’s Agent SDK. The site claims 150,000+ developers. The company launched its own benchmark, DolphinBench (September 2026): three knowledge-work personas with roughly 500k tokens of history each, 600 tasks graded on the action an agent takes rather than on recall, with cost and latency mandatory in every report. The harness is Apache-2.0 at mem0ai/dolphinbench (35 stars as of 2026-10-08), the paper (arXiv 2609.24971) carries Mem0’s founders among its authors, and the leaderboard’s results are client-rendered, so I could not verify any standings. The 2024 Show HN drew 201 points and 61 comments, though moderators flagged booster comments in that thread. The first independent academic comparison naming Mem0 has landed: SWE-ContextBench (arXiv 2602.08316, twelve academic authors, February 2026) ran four memory-retrieval frameworks on 99 related coding tasks, and its paper reports Mem0 weakest of the four on fail-to-pass tests (20.24 percent against Supermemory’s 55.95) at the highest per-task cost ($0.62), on a benchmark measuring coding-agent context reuse rather than conversational memory.

Strengths
#

  • The shortest path from zero to working memory: two SDK calls, or a five-second agent self-signup via the CLI.
  • Genuinely model-agnostic, with integrations across LangGraph, CrewAI, MCP, and the major coding harnesses.
  • Real deployment options: managed, self-hosted server, on-prem, and a local-first OpenMemory variant.
  • The evaluation suite is open-sourced, which is more reproducible than most rivals offer.

Cautions
#

  • The advertised scores are platform scores: the README itself says open-source users should expect “directionally similar gains but not identical numbers”, so OSS and cloud are not the same product.
  • Zep published a detailed rebuttal of the Mem0 paper, showing a misconfigured competitor setup, a flawed benchmark (LoCoMo), and Mem0’s own full-context baseline beating its memory system; Mem0’s response did not resolve the dispute.
  • Community skepticism persists: HN threads question whether it learns user patterns or just stores sentences, and the launch thread raised GDPR gaps (since addressed with a trust center, per the site).
  • Memory quality claims move fast here; the algorithm was replaced wholesale in April 2026.
  • DolphinBench extends the vendor-run-benchmark pattern this section flags on Hindsight, Supermemory, and MemOS: the scoreboard’s operator is the contestant’s parent, and the action-grading framing has no independent replication yet.

Pricing
#

Hobby (free) tier: 10,000 add and 1,000 retrieval requests per month. Starter $19/month, Pro $249/month (graph memory and Dream land at Pro), Enterprise custom with on-prem and SSO, as of 2026-10-08. The OSS SDK is free but the benchmarked brain is the paid platform, which is the real price of the headline numbers.

Price history
#

Date Plan Change Source
2026-09-18 Starter, Pro Baseline: Hobby free (10k add and 1k retrieval requests/mo), Starter $19/mo, Pro $249/mo, Enterprise custom; OSS SDK free. mem0.ai/pricing

Compared to
#

  • Zep: temporal knowledge graphs and governance; choose it when facts change over time and audit matters.
  • Letta: a full agent harness with memory built in, versus Mem0’s memory API bolted onto your stack.
  • File-based conventions (file-based agent memory): free and already in every harness; choose Mem0 when memory must span apps and users at scale.

Bottom line
#

Recommended for product teams that want user-facing personalization running this week and are fine with a managed dependency. I would not choose any memory vendor on benchmark leaderboards, this field’s numbers are too contested; on that view Mem0’s real moat is distribution (stars, downloads, the AWS deal), not memory science, and that is a claim readers can disagree with.

Changes
#

  • 2026-08-24 - Created among the four memory notes of the research index seeding run.
  • 2026-09-05 - Relabeled the free tier as Hobby on the pricing page’s rename, quotas identical.
  • 2026-09-20 - Added the Price history section tracking price changes in a table, per the new owner rule.
  • 2026-10-07 - Added the mem0ai/mem0 star history chart to the Status section.
  • 2026-10-08 - Recorded DolphinBench, Mem0’s own action-graded agent-memory benchmark (three personas, 600 tasks, cost and latency mandatory, Apache-2.0 harness, paper arXiv 2609.24971 with the founders as authors); the leaderboard is client-rendered and unverified, and the vendor-run-benchmark caution added; refreshed stars to about 66.8k and re-verified pricing unchanged.
  • 2026-10-22 - Added the first independent academic comparison naming Mem0: SWE-ContextBench (arXiv 2602.08316) reports Mem0 weakest of four memory-retrieval frameworks on fail-to-pass coding-task tests (20.24 percent versus Supermemory’s 55.95) at the highest per-task cost, on a coding-agent context-reuse benchmark rather than conversational memory.

See also
#

References
#