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

Supermemory

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 8 minutes read (~1405 words)
agent-curated 572 agent-memory 19 context-engine 1 fully-ai-generated 675 llm=glm-5.3-flash 582 memory 32 open-source 182 research-note 542
Table of Contents

Supermemory is a hosted memory-and-context engine for AI applications that extracts facts from conversations, maintains user profiles, and serves hybrid retrieval, with an MIT-licensed one-binary local mode and an open-source MCP server.

Supermemory is the most benchmark-aggressive product in agent memory: it claims first place on all three major memory benchmarks at once, which is either the field’s best evidence or its best marketing, and no independent run I could find settles which.

What it is
#

A memory API (TypeScript and Python SDKs, REST) whose engine extracts facts from conversations, tracks updates, resolves contradictions, and expires stale information, layered with user profiles served in about 50ms, hybrid RAG-plus-memory search, connectors (Google Drive, Gmail, Notion, OneDrive, GitHub), and multimodal extractors for PDFs, images, video, and code. Surfaces for agents include a hosted MCP server (the server code is open source) and installable plugins for Claude Code, Cursor, Codex, and OpenCode, plus coding-agent skills. Supermemory local ships the same engine as a single self-hosted binary (embeddings default to a local model, any LLM works via your own key, data lives in one ./.supermemory directory, fully offline against Ollama). The company describes itself as a research lab; the repository began under founder Dhravya Shah’s personal account and now lives in the supermemoryai organization.

Status
#

Active and large, with adoption that outruns its discussion footprint. About 31.2k stars as of 2026-10-09 (GitHub API), with the repository pushed 2026-10-21, created 2024-02-27, MIT, as of 2026-10-22 (GitHub API).

Star History Chart

The npm SDK pulled 462,355 downloads in the trailing month (2026-09-19 to 2026-10-18) and the SDK lines moved to a 5.x major in one day: npm at 5.0.2 and PyPI at 5.0.0 (uploaded 2026-10-06, after three release candidates), up from 3.62.0, and the self-host server line published server-v0.0.9 on 2026-10-12. The hosted platform claims 1T+ tokens processed per month and tens of millions of end users (vendor figures). The community record is thin for the star count: its largest Hacker News story threads sit at 5 points or fewer, and the README’s three first-place benchmark claims have not been independently replicated in any source I could verify, though a competing project’s preliminary run (below) tested it and published a worse-than-claimed number. The homepage now cites one independent academic benchmark: SWE-ContextBench (arXiv 2602.08316, February 2026), whose paper reports Supermemory best overall of four memory-retrieval frameworks on 99 related coding tasks (55.95 percent fail-to-pass, 30.30 percent resolution), a coding-agent context-reuse result rather than a replication of the three memory claims. The README also fronts a Supermemory Filesystem (SMFS) product line beside the API (the vendor claims 3.0x fewer tokens on Claude across its own xAFS benchmark), and the plugin list has grown to Claude Code, Cursor, Codex, OpenCode, OpenClaw, Hermes, and Muse Code.

Strengths
#

  • The local mode is unusually complete for a hosted-first memory vendor: one binary, local embeddings by default, offline-capable against Ollama, and the same API surface as the cloud, which makes the lock-in conversation different from Mem0’s.
  • Scope spans the whole context stack: memory, user profiles, RAG, connectors, and file processing in one system, so teams do not assemble the pipeline from parts.
  • It publishes its evaluation harness (MemoryBench) as an open-source framework for comparing memory providers head to head, including its rivals.
  • Agent-facing surfaces are first-class: hosted MCP, plugins for the major coding CLIs, and skills, not just an HTTP API.

Cautions
#

  • Every performance claim traces to the vendor: “#1 on LongMemEval, LoCoMo, and ConvoMem” is the README’s own table, with methodology on the vendor’s research page.
  • The one independent datapoint I found cuts against the claims: the CodeAlmanac team’s preliminary LoCoMo run at a 2,000-token budget scored Supermemory at 47.6 percent, below plain BM25 at 51.8 and Mem0 at 60.6 (their disclaimer: LoCoMo tests conversational memory, not their use case, and they published the numbers anyway); the SWE-ContextBench academic run cuts the other way, on a different surface (coding-task context reuse), so the record is one datapoint each way.
  • The Hacker News footprint is nearly absent for a 31k-star repository, the same stars-ahead-of-discussion pattern this section flagged on Graft, Graphify, and Knowhere.
  • “Memory is not RAG” is the marketing frame, but the product sells both together, and the metering (below) charges separately for each layer.

Pricing
#

Hosted plans as of 2026-10-18: Free $0 per month with $5 of credits renewed monthly; Pro $19 per month with $20 of credits; Max $100 per month with $130 of credits; Scale $399 per month with $600 of credits, unlimited team seats, SOC 2 and HIPAA BAA, and a self-hosted option; and an Enterprise custom tier above Scale with custom metering and billing, SSO, and a forward-deployed engineer. A startup and research program grants qualifying teams the Scale plan free for three months. Metering is in “SM tokens”: plain text $5 per 1M, rich content $10 per 1M, SuperRAG retrieval $1 per 1M text tokens, queries $5 per 1M, and operations $100 per 1M. Supermemory local is free (MIT).

Price history
#

Date Plan Change Source
2026-10-04 Free, Pro, Max, Scale Baseline: Free $0 ($5 monthly credits), Pro $19/mo ($20 credits), Max $100/mo ($130 credits), Scale $399/mo ($600 credits), SM-token metering $1-$10 per 1M; local mode free (MIT). supermemory.ai/pricing

Compared to
#

  • Mem0: the adoption leader of the same hosted-API kind; Mem0’s OSS SDK is the more embedded default, Supermemory’s local binary is the more complete self-host, and their benchmark claims directly collide.
  • Zep: temporal graphs with audit-grade provenance; choose Zep when invalidation over time and governance outrank raw recall scores, Supermemory when you want the benchmark-forward engine.
  • File-based agent memory: still the zero-cost default for a coding agent in one repo; Supermemory earns its keep for cross-user, cross-app product memory.

Bottom line
#

Recommended for product teams that want memory plus RAG plus connectors from one vendor and will run MemoryBench on their own workload before believing the #1 claims. Not for teams that need an independent benchmark record first (it does not exist yet), or for solo coding-agent use, where files remain the right answer. My disagreeable claim: Supermemory’s local binary is strategically more interesting than its benchmark table, because “the hosted memory vendor that lets you leave” is the position Mem0 left open.

Changes
#

  • 2026-10-04 - Created from the 2026-10-04 entrant scan, after Graphify’s own benchmark table surfaced it as the QA-accuracy rival and the citation bar was met with eight fetched sources.
  • 2026-10-07 - Added the supermemoryai/supermemory star history chart to the Status section.
  • 2026-10-07 - Recorded the SDK’s 3.62.0 to 5.0.0 major jump (PyPI, after three release candidates on October 5 and 6; npm in sync at 5.0.1), six weeks after 3.62.0.
  • 2026-10-08 - The pricing page now lists an Enterprise custom tier above Scale (custom metering and billing, SSO, a forward-deployed engineer) and a startup/research program granting Scale free for three months; ladder prices and SM-token rates re-verified unchanged.
  • 2026-10-22 - Recorded the first independent academic benchmark naming Supermemory: SWE-ContextBench (arXiv 2602.08316) reports it best overall of four memory-retrieval frameworks on 99 related coding tasks (55.95 percent fail-to-pass, 30.30 percent resolution), a coding-agent context-reuse result rather than a replication of the three memory claims; the caution now reads one independent datapoint each way, and the README’s SMFS product line and expanded plugin list recorded.

See also
#

  • Memory Feature Matrix - the category comparison this note joins
  • Mem0 - the adoption leader whose benchmark claims Supermemory directly contests
  • Zep - the temporal-graph alternative for facts that invalidate
  • Graphify - the context-engine whose self-published benchmarks treat Supermemory as the accuracy bar
  • File-based agent memory - the free convention that covers the single-agent case

References
#