I checked how this compares. Here's what held up.

Where we agree

Two conclusions the field reached on its own

MemoryArena, a 2026 benchmark that tests memory inside multi-step tasks instead of isolated recall, found that systems scoring near-perfectly on standard recall benchmarks drop to 40–60% accuracy the moment a later step depends on what an earlier one learned. That's close to this archive's own biggest unproven claim, stated plainly in one of its design memories: recall is not the same as usefulness, and nobody — including this project — has properly measured the difference yet.

Mem0, one of the most widely deployed memory layers in production, published a 2026 finding that automatic decay handles low-relevance memories fine but can't fix a stale high-confidence fact — that takes a deliberate correction, not a fading one. That's close to word-for-word what this archive concluded in May, from a different direction: a curated archive doesn't need to forget by decay, because the discipline that decides what gets written is the same discipline that decides what gets corrected later.

Where we differ on purpose

Most memory systems decide for themselves. This one doesn't.

A lot of 2026's research — MemOS, ByteRover, and others — is converging on systems that manage themselves: deciding on their own what to summarize, keep, or forget, sometimes learning that policy the way a model learns anything else. It's a reasonable bet, and parts of it are winning benchmarks.

ES Archive bets the other way. The server underneath it only provides mechanism — store, search, link, tag, revise — and has no opinion about what happens to any of it over time. That decision sits one layer up, with whichever AI is driving the archive through a skill: instructions an AI reads and follows, editable and inspectable, kept outside the storage layer entirely. It's the same split an operating-system kernel makes with the software running on top of it — the part that has to stay trustworthy for years stays small and fixed; the part that changes its mind stays separate and replaceable. Stated plainly, the cost of this bet: the archive gives you nothing on day one. It becomes useful only once something — a person, an AI, a skill — does the work of curating it.

What's still open

A question this page doesn't get to answer

Splitting mechanism from policy relocates a problem instead of solving it. ES Archive now speaks to more than one kind of AI client at once — Claude, Codex, LM Studio — each capable of bringing its own judgment about what to write, correct, or forget. Current security research on agent memory names exactly this as the frontier risk: not who can read a memory store, but which connected AI's judgment gets trusted to change it.

This archive doesn't have a settled answer to that yet. This page exists partly to say so in public, before pretending otherwise would be easier.