StartupSeed

Supermemory

Memory and context engine for AI, offered as an API, plugins and a local binary.

Supermemory builds a memory and context layer for AI applications and agents. It extracts facts from conversations, handles temporal changes, contradictions and automatic forgetting, maintains user profiles (about 50 ms per call), and combines RAG with memory in hybrid search. It offers connectors for Google Drive, Gmail, Notion, OneDrive and GitHub, and multimodal extractors. It ships plugins and an MCP server that give Claude Code, Cursor, Codex and OpenCode persistent memory, plus a developer API and a self-hostable single binary. TechCrunch reported a $2.6M seed round in October 2025 for founder Dhravya Shah.

Visit supermemory.ai View on GitHub

Pros & cons

Each point cites its sources and says what kind of evidence backs it. A second reviewer tried to refute every claim before it was kept.

Pros

MetadataMIT-licensed with a very active community (about 31.5k GitHub stars), and the one-binary 'Supermemory local' can run offline with Ollama.[1]
DocsReady-made plugins and an MCP server give Claude Code, Cursor, Codex and OpenCode persistent memory.[1]
User reportsTeams use it so new agent sessions recall decisions made earlier.[1]
DocsLow entry cost: a free plan with $5 of credits a month and a $19/month Pro plan, billed against a published rate card.[1]

Cons

Independent testIts claim to be first on LongMemEval, LoCoMo and ConvoMem is self-reported. In a third party's small LoCoMo test at a 2k-token budget, Supermemory scored 47.6%, below BM25 (51.8%) and Mem0 (60.6%).[1][2]
DocsThe self-hosted option, SOC 2 and the HIPAA BAA start at the $399/month Scale plan, and air-gapped self-hosting is Enterprise-only.[1]
DocsThe local server still uses a hosted reader service to ingest URLs, so it is not fully offline.[1]
DocsUsage needs a positive credit balance, and subscription credits reset every month.[1]

Best for: Developers who want a quick, MCP-ready memory layer for coding agents and chat apps, starting on a cheap tier.
Watch out for: Benchmark retrieval on your own data rather than trusting the self-reported leaderboard results.

Assessed Oct 29, 2026 · adversarially reviewedSuggest a correction

People

Dhravya Shah

Founder & CEO

Context engineering & memory

In research (1)

Learning to Remember: Advances in Memory Mechanisms for Large Language ModelsVerdict: Incremental
Oct 26, 2026 · A commercial agent-memory layer in the same nonparametric-memory family.

San Francisco, CA · Est. 2022

LlamaIndex

The data framework for connecting agents to your data.

LlamaIndex provides ingestion, indexing and retrieval primitives that give agents grounded context over enterprise data — a foundational piece of the retrieval side of the harness stack.

4 pros · 3 cons

Berkeley, CA · Est. 2024

Letta

Stateful agents with memory, from the team behind MemGPT.

Letta (formerly MemGPT, out of UC Berkeley) builds an agent server with built-in memory, context management and state — letting agents persist, learn and be edited like files rather than living in a chat window.

$10M seed led by Felicis (2024)

4 pros · 3 cons

San Francisco, CA · Est. 2021

Exa

Search API purpose-built for AI agents.

Exa (formerly Metaphor) provides neural search and web-crawl APIs tuned for agent consumption — live, cited knowledge for agents instead of stale training data.

4 pros · 4 cons

Connections

Relations from the Harness Hub knowledge graph, as charted in the State of Harness reports.

Built on · 3

Sources

Verified Oct 29, 2026