supermemory vs innernet

supermemory is racing to make retrieval fast and accurate. innernet is trying to make the conclusion right.

Those sound like the same race. They aren't. supermemory's public numbers are about speed and recall — sub-300ms retrieval, "10× faster than Zep, 25× faster than Mem0", state of the art on the standard memory benchmarks. That's genuinely strong engineering and we don't dispute it. But every one of those numbers measures how well a system finds the thing you already gave it.

innernet is built for the part that comes after: what a system makes of what you gave it, and whether it still holds six months later.

A context cloud. Its own words: "the memory layer for AI agents. Context engineering platform powering enterprise APIs, developer plugins, and a personal app that remembers everything." Knowledge-graph memory rather than a plain vector store, SuperRAG for low-latency retrieval, POSIX filesystem mounts, extractors for PDFs, pages, images and audio, connectors for Notion, Drive, Gmail, GitHub and S3, TypeScript and Python SDKs, and a consumer app with a browser extension on top.

It is one of the most technically credible things in this category, and it has the broadest ingestion surface of anyone we compare against. Take that seriously.

A personal memory layer. One command — npx innernet — connects the AI tools already on your machine to a single memory. netti, an agent that is one per person, files what accumulates into Context Maps: named dimensions that read like documents, nodes for the people and decisions that recur, and a commit history you can branch and roll back.

The promise on the front page is "one url. every ai you already use reads and writes the same memory." The promise underneath it is that the memory is about you, not about your documents.

supermemory's unit is a piece of content — a document, a page, a chunk with an embedding. Everything downstream follows from that: ingest more, index better, retrieve faster. innernet's unit is a fact or a decision about a person and their work, with sources attached and a validity window. You can't index your way to "they moved off Postgres in June and here's the reasoning" — something has to have read the work and concluded it.

"Remembers everything" is a real product promise and supermemory delivers on it. But a memory that holds everything equally is a search index, and a search index doesn't know that the second thing you said contradicted the first, or that the contradiction was the interesting part. innernet is deliberately opinionated about what earns a place, because a life is mostly noise and the value is in the compression.

Retrieval systems resolve toward the current best answer. innernet treats the old answer as evidence rather than error — versioned, branchable, still readable. If you want to know when someone's thinking turned, you need the version they turned away from, and most memory systems have quietly deleted it.

• ·Ingestion breadth. PDFs, images, audio, five first-class connectors, real filesystem mounts. We're narrower and we know it.
• ·Published performance work. They benchmark, they publish, they name their competitors' numbers. That's the right behaviour and it's currently ahead of ours.
• ·Developer platform depth. SDKs, self-hosting, a legible pricing ladder from free to enterprise.
• ·Speed. Sub-300ms retrieval is not a thing to hand-wave past.

• ·We're building a picture of a person, not an index of their files. Everything else follows from that.
• ·Disclosure lives on the fact. Every fact carries how far it may travel, enforced when a tool reads. Connecting an AI is not the same as surrendering to it.
• ·Version history you can work inside. Branch a line of thinking, park it, come back and merge it — memory that versions like code.
• ·netti is one per person. Not a shared retrieval service with your namespace inside it, but something that grows into the shape of your work.

Use supermemory if your problem is retrieval — a lot of documents, an agent that needs the right chunk fast, an app that needs a memory API behind it.

Use innernet if your problem is continuity — you want the tools you already use to understand your work, your decisions and how you think, and to keep understanding it as it changes.

"They're state of the art on the benchmarks. Where are your numbers?"

Fair, and we'll publish. But note what those benchmarks measure: whether a system can answer questions about a conversation it was given. That's a real capability and supermemory is excellent at it. It is not a measure of whether a system understood the person in that conversation, and there is no public benchmark for that yet — which is exactly why the ones that exist get quoted so often.

Accurate as of 3 September 2026, based on supermemory's public site. Their performance claims are theirs and we haven't independently reproduced them. If we've described anything unfairly, tell us and we'll correct it.