Kaineros · the landscape

Where this sits

A crowded field, mostly solving different problems.

Plenty of people are building AI memory. Most are answering a different question than it first appears.


The map

Three things get called “AI memory”.

Only one of them is competing with us.

The field, August 2026
KindWhoWhat you get
Memory librariesfor developers Mem0, Zep / Graphiti, Letta, Cognee, LangMem Infrastructure for building agents. An API and a store — you bring the product.
Local-first assistantsself-hostable Khoj, OpenClaw, note-app copilots Answers retrieved from what you already wrote or sent. Very good at finding things. The memory is your existing material.
Hosted assistantscloud, funded Town, and a wave of “personal AI twin” startups Something that genuinely learns you — on their servers, under their terms.
Kainerosthis project An assistant that writes its own memory of you, as plain files on your machine, and decides which version of a fact to believe.

A snapshot, and it will date. Corrections welcome at contact@kaineros.com.

Credit where it’s due

What the others get right.

Serious systems by serious people, several further along than this one. If what you want is on this list, go and use them.

The difference

Three things not on that list.

1. Facts compete. They aren’t overwritten.

Most systems settle a contradiction by recency, or ask a model to merge the two into something agreeable. Kaineros holds the rival versions side by side and ranks them pair by pair, with the deep model as judge. A passing remark has to win to overturn something settled; a deliberate correction takes hold at once.

2. The memory is the artifact, not an index.

Elsewhere it lives in a vector store or a graph you can’t read. Here it’s one Markdown page per fact, with its provenance, in a folder you own. Nothing to export, because nothing was ever encoded.

3. Two speeds, with a latency budget.

Plenty of systems extract memories asynchronously. Here the split is the design: a slow model compiling between conversations, a fast one that only ever reads finished pages — so the half you talk to never pays for the half that thinks.

The other side of it

Where they are ahead.

A comparison that only flatters its author is worth nothing.

  • Better measured. Letta and Zep have been through the standard long-memory evaluations far more thoroughly. What we measured is on the results page — our own corpus, a different benchmark.
  • More places. Khoj runs everywhere you work; OpenClaw answers in whichever app you live in. This is one desktop app.
  • Teams, with funding. This is an early prototype, and the repository is still private.
  • Selection is a bet, and bets can lose. Ranking rival facts costs compute that overwriting doesn’t. If recency turns out to be good enough for almost everyone, the machinery won’t have earned its keep.

A note on numbers

Why there’s no scoreboard.

The rankings you’ll find online trace back to blog posts and to each project’s own marketing, measured on different benchmarks and corpora. Reprinting them would imply a like-for-like comparison nobody has run. So this page compares approaches, which can be described honestly — and the only numbers on this site are ones we measured ourselves, with their method attached.

Same principle as the assistant: say what’s grounded, abstain from the rest.


Get involved

Working on the same thing?

If you’re building in this space — or you think this map is wrong — I’d like to hear from you. contact@kaineros.com

← back to kaineros.com