Both are MCP-native. Both target Claude Desktop and Cursor users. Both claim strong retrieval benchmarks. The architecture and business model are very different. Here's the breakdown.
VEKTOR runs entirely on your machine. The MAGMA graph stores memories across 4 typed layers — semantic similarity, causal relationships, temporal sequences, and entity co-occurrence. AUDN evaluates every new input before writing: ADD, UPDATE, DELETE, or NO_OP. Zero duplicates by design.
The REM compression cycle runs asynchronously — 50 fragments become 3 distilled insights while your agent sleeps. All retrieval is local SQLite: 8ms average, no API roundtrip, no embedding bill beyond your existing LLM key.
Supermemory provides a unified cloud API for fact extraction, user profile building, contradiction resolution, and selective forgetting. It wraps retrieval complexity behind a single remember()/search() interface.
The internal architecture is not publicly documented — it's a managed black-box API. Supermemory claims benchmark leadership on LongMemEval, LoCoMo, and ConvoMem, but as of May 2026 these scores have not been independently verified by third parties. The self-reported numbers may reflect favourable evaluation conditions.
The browser extension is a genuine differentiator: it lets agents capture personal knowledge from web browsing, which has no equivalent in VEKTOR.
Supermemory leads its positioning with benchmark scores on LongMemEval and LoCoMo. These are real benchmarks — but the scores are self-reported and, as of this writing, haven't been reproduced by independent evaluators. This is common in the memory layer space (several competitors have made similar claims later shown to be architecture-specific rather than general).
VEKTOR's published metrics — 8ms recall, 97.3% precision, 50:1 REM compression — are internal production figures from the VEKTOR runtime, not benchmark suite scores. They're measuring different things. Neither set of numbers is directly comparable to the other.
The practical question isn't which number is bigger — it's whether retrieval is fast and accurate enough for your agent's workflow. A 5-minute install on both will tell you more than any benchmark.
VEKTOR is $9/month flat. Supermemory offers a free tier, a Pro tier, and Enterprise — exact Pro pricing isn't publicly listed. For teams running agents at any meaningful query volume, usage-based cloud billing tends to compound faster than expected. VEKTOR's flat model eliminates this unpredictability entirely.
These are the two most MCP-native memory layers in the market right now — and they're genuinely different products. Supermemory wins on browser extension, managed cloud, and multi-user scenarios. VEKTOR wins on latency, pricing predictability, data ownership, architectural transparency, and Node.js-native depth. If you're building a Claude Desktop or Cursor workflow that needs local-first memory you own completely, VEKTOR is the stronger fit. If you need managed cloud with a browser knowledge capture layer, Supermemory is worth evaluating.
Local-first. 8ms recall. MCP-native. $9/month flat.