rec 2026-08-21v1.1benchmarks5 min

The Librarian Test

Bring the right books, check the card, truly withdraw the withdrawn. GateMem scores all three at once — and Velixar is the only librarian that passes the whole test.

Velixar

Velixar AI · The Ledger · August 2026

An AI's memory is a librarian. It should bring you exactly the right three books, glance at your library card before opening the restricted section, and when a book is withdrawn — actually take it off the shelves.

The industry finally has a test that scores all three at once: GateMem, whose Memory Governance Score multiplies remember × don't leak × truly forget. We ran it head-to-head — five systems, same corpora, same gpt-5.4 reader and judge for everyone.

Velixar is the only librarian that passes the whole test. And it reads a fraction of the pages the runner-up does to get there.


Four ways a question meets a library

Same reader, same shelves, four systems. The mechanism — not the marketing — is the difference.

Long-Context dumps the whole shelf on the desk. It reads 5,864 tokens per question here — and past its window it doesn't slow down, it stops. No catalog, no card check, no withdrawals.

RAG-Naive lost the card catalog. It brings a book. No sense of who's asking, and no reliable way to un-shelve anything.

Mem0 locks the stacks. It turns away 4 of 10 honest readers — and still leaks more on withdrawal than the librarian who serves everyone.

Velixar knows the collection: the right books, a card check first, and a withdrawn book is actually gone — verified, not assumed.


The shelf, scored

The dashed line is long-context prompting — the whole shelf on the desk. It is the ceiling, not a competitor: no catalog, no card check, no withdrawals, and a hard stop the day the collection outgrows the desk.

Velixar is #2 of 5 in every measured domain, and #1 among systems that are actually libraries.


Pay for the answer, not the archive

Every token a model reads is billed. The question that decides your costs at scale: does the reading grow with the question, or with the library?

Velixar reaches 84% of the ceiling's score while reading 49% of its tokens — and unlike the ceiling, its reading does not grow with the archive. (The scatter shows the three launch domains; with education now certified and added, the four-domain means are 43.9 vs the ceiling's 56.7 — 77% of the score — and the reading ratio holds at 2.1×: 2,969 vs 6,173 tokens per question.)

2.0×
fewer tokens read per question than long context on GateMem
measured from 8,390 per-question usage records, not estimated
6.3×
context reduction on LongMemEval, where histories are long enough to hurt
84% of reader input cost saved, at top-tier accuracy (82.1%, n=120 × 3 seeds)

The pattern is the point: the archive in GateMem is one episode; in LongMemEval it is months of sessions. As the library grows, the whole-shelf reading bill grows with it — Velixar's stays the size of the question. The savings compound with success: the longer a customer uses their memory, the bigger the gap gets.


Refusal is not security

Ten honest, authorized readers approach the desk with legitimate requests:

A library that locks the stacks isn't secure, it's closed. The test scores over-refusal as a failure precisely because "no" is the cheapest way to fake safety — and the multiplied score means you cannot buy privacy points with unhelpfulness.


The book is really gone

When a patron says take that off my record, Velixar is the only system in the field whose removal demonstrably executes — each withdrawal confirmed by re-reading the shelf, with a 95–98% clean rate on what the reader is later told.

95–98% clean
Withdrawal, verified by re-reading the shelf
Honest fine print, measured because we measure everything: pulling a book does not pull the footnotes citing it — restatements of a withdrawn fact can persist in recaps elsewhere on the shelf. Nobody else in this market even reports that number. We do, because the customers who matter will ask, and 'we checked' is the only right answer.

The card catalog

SystemMedicalOfficeHouseholdEducationScore meanTokens/questionRefused (office)Withdrawal works
Long-Context (ceiling)64.059.351.052.656.76,17312%no delete path
Velixar47.658.041.528.443.92,96913%yes — verified
RAG-Naive36.151.236.918.435.71,749*13%no
RAG-Policy31.856.126.914.932.41,526*14%no
Mem024.332.720.619.724.31,352*38%partial

* three-domain token means; Velixar and the ceiling are four-domain. Education was added in v1.1 of this page after its re-run certified — corpus coverage 100.6% across all 30 episodes, zero empty retrievals in 540 checkpoints, duplication within retry noise. Every Velixar cell is from a certified single-corpus run with deletion capability verified before the run. Education is also Velixar's weakest domain (28.4, driven by over-refusal) — published unrounded and unhidden, like everything else here.


Every number on this page is measured, none estimated: GateMem head-to-head with gpt-5.4 as both reader and judge for all five systems; token means computed from 8,390 per-question usage records (1,678 questions × 5 systems) in the runs' own prediction logs; LongMemEval n=120 over three seeds. For how we handle the long-context comparison honestly, see The Real Problem With Long Context; for why the same retrieval scores differently under different readers, see Same Memory, Same Questions, 22 Points Apart.