Production polish: sub-300ms, SOC2/HIPAA, connectors, context fencing via container tags. Writes auto-extract facts into dual-layer user profiles (static + dynamic) and a unified knowledge graph; an asynchronous background pass ('Dynamic Dreaming', shipped May 2026) reconsolidates fragmented memories, reweights contradictions, and infers cross-memory connections without blocking the write path. Claims #1 on LongMemEval/LoCoMo/ConvoMem (self-reported); publishes an open-source benchmarking harness, MemoryBench, that others can run against Mem0 and Zep, though Supermemory's own headline numbers still come from its own runs of it.
Deep diveSupermemory deep dive: how the automatic memory API differs from Mem0 in practiceRead →- Storage
- Cloud API, or local binary (full Memory API on localhost via one-command installer, embedded graph engine + local embeddings, BYO model incl. fully offline via Ollama)
- Retrieval
- Managed semantic retrieval with 'context fencing' (container-tag isolation) to keep tenants' memories from polluting each other; hybrid search combines RAG + memory in a single query; one call (~50ms claimed) returns both the ranked profile summary and relevant memories.
- Self-host
- Self-host: trivial
- License
- MIT
- Pricing
- Free ($5/mo incl.) → $19 → $100 → $399 → Enterprise, plus unit-based 'Context Cloud' usage pricing under the included credits (Memory $0.005/1K tokens, SuperRAG $0.001/1K, Search $0.005/1K queries, Operations $0.10/1K) · Freemium
- GitHub stars
- 28,842
- Last release
- 2026-07-19
- Last commit
- 2026-08-10
- First catalogued
- 2026-06-28
Strengths
- Anti-pollution context fencing via container tags
- Sub-300ms
- Multi-tenant
- Local binary self-host option (one command, zero config)
- Connectors: Google Drive, Gmail, Notion, OneDrive, GitHub
- Hybrid search: RAG + memory in one query
- Asynchronous consolidation ('Dynamic Dreaming', May 2026): reconciles contradictions and infers cross-memory connections in the background without blocking writes
- Automatic ingestion model: no manual add()-style memory calls, chunking strategy, or vector DB config required
- Broad framework-adapter surface (Vercel AI SDK, LangChain, LangGraph, Mastra, Agno, CrewAI, Microsoft Agent Framework, Convex, VoltAgent, n8n, Zapier) plus coding-agent plugins for Claude Code, OpenCode, OpenAI Codex CLI, OpenClaw, and Hermes
- Self-hosted local binary now supports pluggable embeddings (local ONNX default, or OpenAI/Google/Ollama) as of server-v0.0.5 (Jul 2026), narrowing the self-host/cloud capability gap
Watch out
- Benchmark claims (#1 LongMemEval/LoCoMo/ConvoMem; SMFS's 55%-cheaper/more-accurate retrieval claim) are self-reported; no independent reproductions on file. Numbers also vary by page and methodology within Supermemory's own materials — its technical report claims 81.6% (GPT-4o) / 84.6% (GPT-5) / 85.2% (Gemini-3 Pro) LongMemEval accuracy via its own MemoryBench harness, while its research page separately claims 95% Recall@15 with aggregation (a different metric) — read any single number against its stated config, not as one settled score. A December 2025 Hindsight paper (arXiv:2512.12818) reprints these same Supermemory-reported figures rather than independently re-running them, and Mem0's own vendor comparison page pits its self-reported 94.4 against Supermemory's self-reported 85.2 as if the two were run under one shared methodology — they weren't
- Automatic, profile-driven extraction trades off the fine-grained manual control that explicit-add frameworks like Mem0 offer; an April 2026 community comparison flagged the lack of first-class consolidation/contradiction-detection prior to Dynamic Dreaming shipping
- SOC 2 + HIPAA BAA require Scale plan ($399+/mo) or Enterprise — lower tiers (Free, Pro, Max) do not include formal compliance certifications
- Full cloud API self-hosting also gated to Scale/Enterprise; local binary may lack multi-tenant features
- Proprietary internals on the cloud tier
Best for
- Polished managed memory API with SOC 2 / HIPAA compliance
- Coding-agent memory via MCP (Claude Code, OpenCode, OpenAI Codex CLI plugins)
- One API over mixed data (files, email, PDFs, chat)
How it integrates
Benchmark results
No sourced results yet.
Sources
- Pricing page re-checked — unchanged: Free $0 (~$5/mo usage included), Pro $19/mo (~$20 usage), Max $100/mo (~$130 usage), Scale $399/mo (~$600 usage), Enterprise custom; Context Cloud unit rates still Memory $0.005/1K SM tokens ($0.010 rich), SuperRAG $0.001/1K ($0.002 rich), Search & Traversal $0.005/1K queries, Operations $0.10/1K (vendor)
- Monthly card re-verification — GitHub API repo metadata: 28,724 stars, license MIT confirmed, archived=false, last commit on the default branch 2026-08-01 (third-party)
- https://supermemory.ai/ (vendor)
- https://github.com/supermemoryai/supermemory (vendor)
- Supermemory — universal memory API; MCP plugins for coding agents (vendor)
- Supermemory self-hosting docs — local binary and Scale-tier API self-hosting (vendor)
- Supermemory pricing — SOC 2 + HIPAA BAA confirmed on Scale ($399/mo) and Enterprise tiers only (vendor)
- Supermemory MCP 4.0 setup docs — remote-http server at https://mcp.supermemory.ai/mcp; OAuth or API key auth (vendor)
- Supermemory REST API reference — base URL https://api.supermemory.ai/v3; ingest, recall, profiles, knowledge graph endpoints (vendor)
- supermemory on PyPI — official Python SDK v3.50.0 (Jun 2026) from supermemoryai/python-sdk (vendor)
- supermemory on npm (registry API) — official TypeScript SDK v4.24.12 (Apache-2.0), repo supermemoryai/sdk-ts, ships type declarations (third-party)
- supermemoryai/supermemory README — local binary architecture (embedded graph engine, local embeddings, BYO model incl. offline Ollama), container-tag context fencing, MCP tools (memory/recall/context), dual-layer static+dynamic user profiles, AST-aware code chunking (vendor)
- Introducing Dynamic Dreaming — asynchronous background pass that reconsolidates fragmented memories, reweights contradictions on newer context, and infers traceable cross-memory connections; runs on inactivity or context buildup, typically completing within ~15 minutes (vendor)
- SMFS (Supermemory Filesystem) — agentic/semantic filesystem for agents; self-reported 55% cost reduction and 81% vs 69% accuracy at 10,000 files on the authors' own xAFS benchmark (smfs.ai/research/runs) (vendor)
- Context Cloud pricing announcement — unit-based pricing (Memory $0.005/1K tokens, SuperRAG $0.001/1K, Search $0.005/1K queries, Operations $0.10/1K) layered under the existing tier credits (vendor)
- Supermemory pricing page — confirms tier prices and unit-based pricing both current as of this date (vendor)
- LogRocket (independent) — Mem0 is explicit/developer-managed memory (you decide what to add/retrieve); Supermemory infers, updates, and retrieves automatically via user profiles, trading manual control for less application logic (third-party)
- Hacker News (comment on 'A memory database that forgets, consolidates, and detects contradiction') — April 2026 community comparison noting Supermemory lacked first-class consolidation ('think()') and contradiction detection as of that date, prior to Dynamic Dreaming shipping in May (third-party)
- GitHub API — supermemoryai/supermemory repo stats: 28,393 stars, pushed 2026-07-15, MIT license (vendor)
- supermemory-server v0.0.5 release notes — pluggable embeddings for self-hosted (local ONNX default, or OpenAI/OpenAI-compatible/Google via env config); safe-upgrade lock file (vendor)
- supermemory-server v0.0.4 release notes — smarter memory generation from documents, smoother ingestion, better file handling, time-aware search, richer /v4/profile buckets, clearer search ranking (vendor)
- Supermemory docs index (llms.txt) — confirms current integrations list: Vercel AI SDK, LangChain, LangGraph, OpenAI Agents SDK, Mastra, Agno, CrewAI, Microsoft Agent Framework, Convex, VoltAgent, n8n, Zapier, viaSocket, plus coding-agent plugins for Claude Code, OpenCode, OpenAI Codex CLI, OpenClaw, and Hermes; connectors add Granola, S3, and Web Crawler beyond the previously-verified set (vendor)
- Supermemory changelog — Supermemory CLI shipped Mar 2026 (manage memories/documents/profiles/tags/connectors/API keys from the terminal); Microsoft Agent Framework Python package (supermemory-agent-framework) shipped Mar 2026; instant-dreaming API parameter shipped May 27 2026 (vendor)
- MemoryBench docs — Supermemory's open-source, self-hostable benchmark harness for memory providers (works with Supermemory, Mem0, Zep); a genuinely re-runnable tool, but Supermemory's own headline scores still come from its own runs of it (vendor)
- Latimer et al., 'Hindsight is 20/20: Building Agent Memory that Retains, Recalls, and Reflects' (arXiv:2512.12818, submitted 2025-12-14) — Table 3 reprints Supermemory's own self-reported LongMemEval figures (81.6% GPT-4o / 84.6% GPT-5 / 85.2% Gemini-3 Pro) 'taken directly from the Supermemory technical report' rather than independently re-running them; Table 4 likewise treats Mem0/Supermemory LoCoMo numbers as 'reported reference points,' not independent reproductions (paper)
- Mem0's own vendor comparison page — cites its self-reported LongMemEval 94.4 (config undisclosed) against Supermemory's self-reported 85.2, states Supermemory has 'not published' LoCoMo/BEAM; vendor framing, not a neutral same-config test (vendor)
- GitHub API license endpoint — root LICENSE file is MIT (name 'MIT License'), confirmed by decoding contents; corrects the prior placeholder 'OSS*' value. Repo also carries per-package LICENSE files (skills/, packages/tools, packages/agent-framework-python, packages/openai-sdk-python) all under the same MIT terms (third-party)
- Supermemory pricing re-check — unchanged: Free $0, Pro $19/mo, Max $100/mo, Scale $399/mo (SOC2 + HIPAA BAA), Enterprise custom (SOC2 + HIPAA + GDPR) (vendor)
- supermemory-server v0.0.6 (2026-07-19) — adds Windows self-hosted server support (vendor)
- GitHub API re-check — not archived, pushed 2026-07-23T19:18Z (third-party)
Last verified 2026-08-01 · updated by refresh-framework-cards