Three public memory benchmarks. Every system answers with the same model
and is graded by the same judge.
79.0
LoCoMo: 5th of 16
81.4
LongMemEval-S: 3rd of 14
71.4
BEAM-100K: 2nd of 13
0
AI calls when Memoria stores a memory
Where it ranks
Gaps of less than about 5 points are noise.
LoCoMo
Memoria 79.0, 5th of 16
accuracy (%), 1,540 questions
LongMemEval-S
Memoria 81.4, 3rd of 14
accuracy (%), 500 questions
BEAM-100K
Memoria 71.4, 2nd of 13
nugget score (%), 400 questions
Accuracy against context size
Context is the text the answer model reads per question: it sets the cost
and delay of every answer. Up and to the left is better.
LoCoMo
Memoria sends 1,369 tokens per question; 14 of 15 send more
LongMemEval-S
Memoria sends 8,085 tokens per question; 5 of 13 send more
BEAM-100K
Memoria sends 19,084 tokens per question; 3 of 12 send more
In real work
610 Memoria searches in 141 of our own
Claude Code sessions, read from the session logs.
93%
of used memories were written before that session started
1 in 3
were more than a month old
How old the used memories wereWhere they came from
A memory counts as used only when the agent's next output repeats an
identifier from it, such as a file name, a number or a ticket key. Paraphrase
is not counted, so these are lower bounds.
With Memoria and without
Internal test, not a benchmark
The same Claude agent answers the same questions about our own work, once
with Memoria search and once without. Most questions ask about knowledge
that is in Memoria and not in the files; the last two rows are controls.
Without MemoriaWith Memoria
Project history, 4 code repos
68 questions
+38 pts
Facts only Memoria holds, notes vault
28 questions
+32 pts
All questions, notes vault
69 questions
+14 pts
Facts already in the vault (control)
18 questions
no change
Code facts in the repo (control)
12 questions
no change
0%25%50%75%100%
Correct answers (%)
+38 pts
correct answers on project-history questions over four code repos: 49% without Memoria, 87% with it
−43%
total prompt tokens over the notes-vault questions: Memoria replaces long file searches that end in "not found"
Our own test, not an independent benchmark. We wrote the questions about our
own notes and four of our own code repositories (no client data). Claude Opus
graded the answers, and it could see which run used Memoria; a human check is
still to come. On the control questions, where the answer is already in the
files, accuracy did not change. This suggests that the gains come from
knowledge the files do not hold. Token change 95% CI −62% to −13%.
How we measured
OmniMemEval, an open harness from MemTensor, unchanged apart from registering Memoria. Other systems' scores are as OmniMemEval publishes them (retrieved 12 November 2026).
Every system: gpt-4.1-mini writes the answers, gpt-4o-mini judges them. Memoria: Postgres store, knowledge graph off, top 20 results.
MemTensor makes MemOS, which leads two of the three benchmarks. The harness and its results are open.