Supermemory

Executive Summary

Supermemory is a memory API and context engine for AI agents — the premise being that agents need to learn and grow with the user rather than start cold every session. Applications push unstructured data (documents, chats, PDFs, emails, images, audio) through a single API; Supermemory's custom vector graph engine builds ontology-aware connections and serves hybrid vector-plus-keyword retrieval at claimed sub-300ms latency, along with deep user profiles inferred from behavior — intent, preferences, and context.^[1][2] It began as an open-source consumer "second brain" app in 2024 before pivoting to B2B memory infrastructure; the MIT-licensed repo has 26.8K stars as of June 2026.^[3][4]

The founder story is the headline: Dhravya Shah was 19 when the seed round was announced in October 2025 — a Mumbai teenager who built bots while preparing for IIT entrance exams, sold a tweet-formatting bot to Hypefury, used the proceeds to move to the US, and dropped out of college to build Supermemory.^[5][6] The round — $2.6M per press, $3M per the company — was led by Susa Ventures with Browder Capital and SF1.vc, plus angels including Google Chief Scientist Jeff Dean, Cloudflare CTO Dane Knecht, DeepMind's Logan Kilpatrick, Sentry founder David Cramer, and Theo Browne.^[6][3] At the raise the company cited 50K+ consumer-app users, enterprise customers including Cluely and Composio, and customers processing billions of tokens weekly.^[3] Tracxn lists $29M across two rounds (including a "$26.0M seed" in October 2025), but no primary source confirms a round beyond the ~$3M seed — the larger figure should be treated as reported, unconfirmed, and possibly a data error.^[7]

---

Product Overview

The integration model is deliberately minimal: add memory, RAG, user profiles, and data connectors to an agent through one API — no vector database to configure, no embedding pipelines, no chunking strategies.^[1][2] Data flows in via REST/OpenAPI endpoints, TypeScript and Python SDKs, or pre-built connectors (Notion, Google Drive, Gmail, S3, web crawler); multi-format extractors handle PDFs, web pages, images, and audio.^[1][8] On the retrieval side, the platform combines vector and keyword search over a custom vector graph with ontology-aware connections, and builds user profiles from accumulated behavior so agents carry intent and preferences across sessions.^[1]

Key Capabilities

Product Surfaces

---

Technical Architecture

Supermemory is a hosted-first service with an open-source escape hatch: the GitHub repo describes a "memory and context engine + app that is extremely fast, scalable, and can be run fully locally."^[4] The company says it built its own vector database, content parser, and extraction tooling rather than assembling off-the-shelf RAG components, optimizing for retrieval latency at scale.^[3] The website claims sub-300ms retrieval latency on hybrid search.^[1]

Key Technical Details

---

Strengths

• Credibility-dense cap table for a tiny round. Jeff Dean, Cloudflare's CTO, DeepMind's Logan Kilpatrick, and Sentry's founder backing a ~$3M seed is an unusual signal-to-dollars ratio, and the angels map directly to the infrastructure problem being solved.^[6][3]
• Real open-source gravity. 26.8K stars, 2.3K forks, MIT license, and an actively pushed repo (latest commits June 2026) — the project predates the company pivot and gives developers a fully local fallback.^[4]
• Ingestion breadth as differentiation. Multi-format extractors (PDF, web, image, audio) plus managed connectors mean teams feed it raw organizational exhaust rather than pre-cleaned text — closer to a context pipeline than a key-value memory store.^[1]
• Production traction beyond stars. Named customers Cluely and Composio, 50K+ consumer-app users, and customers processing billions of tokens weekly at the time of the raise.^[3]
• Generous on-ramp. Free tier includes $5/month of usage plus the MCP server and Hermes plugin, so evaluation costs nothing.^[8]

---

Cautions

• Funding data is muddled. Tracxn lists $29M across two rounds including a "$26.0M seed" — suspiciously a 10x of the confirmed $2.6M, and no press release, blog post, or primary source corroborates it. Treat anything beyond the ~$3M seed as reported, unconfirmed.^[7][6]
• Solo, very young founder. Shah's trajectory is impressive, but a single 20-year-old founder carrying infrastructure SLAs for enterprise customers is concentration risk that buyers should price in.^[5]
• Benchmark claims are vendor-stated. "Tops every benchmark" and sub-300ms latency come from the company's own blog and website; independent benchmark validation is scarce, and this category has a track record of disputed benchmarks (Mem0 vs. Letta vs. Zep).^[3][1]
• Crowded, churning category. Memory layers are proliferating faster than they are consolidating; developers are already building backend-agnostic abstraction layers specifically to swap providers "as they come and go on the leaderboards."^[9][10]
• No standalone breakout community moment. Despite the star count, Supermemory has no high-engagement dedicated HN launch thread; community discussion appears mostly in passing within broader agent-memory debates.^[10]

---

What Developers Say

Community discussion as of June 2026 is real but diffuse — Supermemory surfaces in agent-memory threads rather than headlining its own.

"We also implemented supermemory.ai, so newly made decisions are always recalled by AI agents when starting new sessions." — an HN commenter on production agent workflows^[11]

"I got tired of having to change my code whenever I wanted to try a different memory backend, like mem0, letta, supermemory, hindsight." — an HN commenter who built a backend-agnostic SDK^[9]

"Supermemory is a hosted SaaS built around embedding + ranking." — the author of a competing self-hosted memory project, contrasting it with consolidation and contradiction-detection features^[12]

"There are a quadrillion startups (mem0, langmem, zep, supermemory), open source repos (claude-mem, beads), and tools that do this." — an HN commenter skeptical of the category^[10]

---

Pricing & Licensing

API usage beyond included credits is metered: memory storage at $0.005-0.010 per 1K tokens, SuperRAG at $0.001-0.002 per 1K tokens, search at $0.005 per 1K queries, and operations at $0.10 per 1K.^[8]

Licensing model: MIT open source for the core engine and app; the hosted platform (connectors, team features, compliance) is the proprietary commercial layer.^[4][8]

Hidden costs: usage metering past the built-in credits can outrun the headline tier price for token-heavy agents, and the most useful connectors (Gmail, GitHub, S3) are gated behind the $100 and $399 tiers.^[8]

---

Competitive Positioning

Direct Competitors

When to Choose Supermemory Over Alternatives

• You want one API that handles ingestion (PDFs, audio, email, Drive) and retrieval — not just a memory CRUD layer you feed pre-processed text.^[1]
• Retrieval latency matters and the sub-300ms claim survives your own benchmark.^[1]
• You want an MIT-licensed local fallback if the vendor relationship sours.^[4]
• You need SOC 2/HIPAA or air-gapped deployment without graduating to building memory infrastructure in-house.^[8]

---

Ideal Customer Profile

Best fit:

• Agent and LLM-app teams that want persistent user memory without owning vector-DB, chunking, and embedding infrastructure^[1]
• Products whose context lives in messy formats — email, docs, audio, PDFs — where Supermemory's extractors and connectors replace a custom ingestion pipeline^[1]
• Startups that value the free tier + MIT escape hatch for low-commitment evaluation^[8][4]

Poor fit:

• Enterprises that need a vendor with deep organizational redundancy — this is a young, founder-concentrated company^[5]
• Teams already on LangGraph (LangMem) or wanting a full stateful agent runtime (Letta) rather than a memory API
• Buyers who require independently audited benchmarks before adopting performance claims^[3]

---

Viability Assessment

The angel roster buys Supermemory time and credibility disproportionate to its raise, and the open-source base gives it distribution Mem0 had to build the same way. The open questions are durability ones: a solo young founder, vendor-stated benchmarks, and a memory category that developers themselves describe as having "a quadrillion startups."^[10]

---

Bottom Line

Supermemory is the highest-profile young-founder bet in agent memory: a real product with real open-source traction (26.8K stars), real customers (Cluely, Composio), and an angel list — Jeff Dean, Dane Knecht, Logan Kilpatrick — that functions as technical due diligence by proxy.^[4][3][6] It is earlier and smaller than Mem0, and claims of $29M raised are unconfirmed and likely a database error; judge it as a well-angeled ~$3M seed company.^[7]

Recommended for: agent builders who want memory plus ingestion as one managed API, teams that value an MIT-licensed local fallback, and free-tier evaluators.

Not recommended for: enterprises needing organizational depth behind SLAs, LangGraph-committed stacks, or buyers requiring independent benchmark validation.

Outlook: strong seed-stage signal in a brutally crowded category — the next 12 months of enterprise logos and any confirmed follow-on round will determine whether it stays a challenger or becomes the consolidation target.

---

Research by Ry Walker Research • methodology