Retrieved, Not Read

What we actually sell

Velixar is a retrieval and memory layer. It is not a reader, and the model that consumes what we return is the customer's choice — swapped freely, priced independently, and improving on a schedule none of us control.

So the number that describes our product is the one that measures retrieval on its own. In Measuring Memory we showed that an end-to-end benchmark score is a joint measurement of a memory system and a reader, and that changing only the reader moved the score by 16.4 points. This paper reports the half of that measurement that is ours.

Setup

LoCoMo (Maharana et al.) — 10 long multi-session conversations, 272 sessions, 1,986 questions. We ran a 400-question stratified sample, category-proportional to the full corpus, selected by fixed seed so every arm answers identical questions. Retrieval budget k=25. Isolation enforced per conversation and asserted at answer time.

LoCoMo labels the evidence session behind each question, which makes retrieval scoring deterministic: did the gold session come back, or not. No reader, no judge, no prompt.

98.3%

This table is the canonical recall curve for our published results; companion papers reference it rather than restate it.

k	recall@k
1	38.2%
3	65.5%
5	78.1%
10	89.6%
20	96.5%
25	98.3%
30	99.9%

Many LoCoMo questions require several evidence sessions, not one. Scoring only questions where every gold session was returned: 97.4% at k=25.

Per category, at k=25: single-hop 98.2%, multi-hop 96.3%, temporal 100%, open-domain 100%.

What memory is worth

Retrieval quality is a property of our layer. Its value only appears against the alternative, so we measured that directly: the same reader, the same questions, with memory and without.

+55.0 pts

This is the one comparison in this paper that holds a reader constant on both sides, which is what makes it attributable. A memory layer that cannot show its lift against the same model answering unaided has not demonstrated it is doing anything.

Retrieval budget and cost

The recall curve is informative beyond its endpoint. Recall reaches 89.6% at k=10 and 99.9% at k=30 — the evidence is retrievable well before the budget we used, but is not always ranked into the top ten.

Context in this corpus runs about 553 tokens per session, so k=25 costs roughly 13,800 tokens per query and k=10 roughly 5,500. Retrieval budget is therefore a direct cost lever, and one that is almost never disclosed alongside a published accuracy figure.

Why we report retrieval separately

The gap between retrieval and answers is not rhetorical. On this run, retrieval returned the evidence for 98.3% of questions while end-to-end accuracy on categories 1–4 was 68.9%.

That distance is reader-side, and we can show it three ways. Restricting to questions where the gold session was inside the context the reader saw, accuracy is 70.8% — so position in the context is not the explanation. Ranking the gold session first moves it only to 76.2%. And an earlier arm with identical retrieval and an identical judge, differing only in reader, scored 56.7%.

Twelve points of spread are available from reader choice before any memory system is involved. That is the practical case for reporting the two halves separately: a single blended number moves for reasons that have nothing to do with the layer being evaluated.

A result we withdrew

We initially read our temporal sub-score as a weakness in temporal handling. It is not, and the reason generalizes.

These benchmarks are commonly run by ingesting an entire corpus in a single pass. The conversations in LoCoMo span months; the memories representing them are written within seconds of each other. Whatever ordering the corpus had in the world is absent from the store's time axis and survives only as text inside the content.

A temporal score produced this way measures a model's ability to do date arithmetic on dates written in prose. It does not measure a memory system's temporal handling, because under this protocol that machinery has nothing to act on. Published results do not generally state whether their harness back-dates ingestion — and without it, temporal sub-scores should not be read as a property of the memory system.

We are withdrawing our temporal figure rather than defending it. Measuring this honestly requires back-dated ingestion or a corpus accumulated over real elapsed time.

Reproducibility

Every figure above came through the standard client path against the shipping product, not a tuned configuration. Retrieval budget, sample size, seed, sampling method, and — where a reader is involved — the reader and judge are stated, because a benchmark result that omits them is not checkable.

Three measurement defects in our own harness were found and corrected while producing this paper, including one that would have published a materially wrong figure by silently dropping rows from a result rather than retrying them. Harness correctness is part of the claim.

@techreport{velixar-retrieved-not-read-2026,
  title       = {Retrieved, Not Read},
  author      = {Velixar AI},
  year        = {2026},
  institution = {Velixar},
  url         = {https://velixarai.com/research/retrieved-not-read}
}