How it works Oracle · built on Memoria Vs vendor memory Benchmarks Writing Get access
Theme

Benchmarks

How Memoria measures up

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
±5 pts (noise) 0 25 50 75 100 MemOS: 88.83 MemOS 88.8 Cognee: 83.48 Cognee 83.5 EverOS: 82.75 EverOS 82.8 Hindsight: 81.99 Hindsight 82.0 Memoria: 78.96 Memoria 79.0 LCM-X: 78.83 LCM-X 78.8 Mem0: 77.68 Mem0 77.7 Letta: 77.12 Letta 77.1 MemMachine: 73.90 MemMachine 73.9 mem9: 73.64 mem9 73.6 Supermemory: 73.53 Supermemory 73.5 MemoryLake: 72.49 MemoryLake 72.5 Viking: 69.33 Viking 69.3 Zep / Graphiti: 63.83 Zep / Graphiti 63.8 Memori: 41.34 Memori 41.3 Backboard.io: 22.40 Backboard.io 22.4

accuracy (%), 1,540 questions

LongMemEval-S Memoria 81.4, 3rd of 14
±5 pts (noise) 0 25 50 75 100 MemOS: 89.20 MemOS 89.2 LCM-X: 87.00 LCM-X 87.0 Memoria: 81.40 Memoria 81.4 EverOS: 80.40 EverOS 80.4 Zep / Graphiti: 79.80 Zep / Graphiti 79.8 mem9: 78.00 mem9 78.0 Letta: 77.67 Letta 77.7 Hindsight: 72.20 Hindsight 72.2 Supermemory: 66.07 Supermemory 66.1 MemMachine: 63.60 MemMachine 63.6 Viking: 61.07 Viking 61.1 Mem0: 56.00 Mem0 56.0 Cognee: 51.80 Cognee 51.8 Memori: 20.80 Memori 20.8

accuracy (%), 500 questions

BEAM-100K Memoria 71.4, 2nd of 13
±5 pts (noise) 0 25 50 75 100 LCM-X: 71.84 LCM-X 71.8 Memoria: 71.39 Memoria 71.4 Viking: 70.76 Viking 70.8 Mem0: 70.41 Mem0 70.4 Hindsight: 70.22 Hindsight 70.2 Letta: 69.22 Letta 69.2 Zep / Graphiti: 68.70 Zep / Graphiti 68.7 MemOS: 66.87 MemOS 66.9 Supermemory: 65.98 Supermemory 66.0 mem9: 65.75 mem9 65.8 MemMachine: 64.80 MemMachine 64.8 Cognee: 59.30 Cognee 59.3 EverOS: 58.64 EverOS 58.6

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
50 60 70 80 90 1k 10k 100k context tokens per question (log scale) → ↑ accuracy (%) MemOS: 88.8 at 5,400 tokens Cognee: 83.5 at 32,532 tokens EverOS: 82.8 at 8,559 tokens Hindsight: 82.0 at 24,683 tokens LCM-X: 78.8 at 1,659 tokens Mem0: 77.7 at 17,395 tokens Letta: 77.1 at 14,188 tokens MemMachine: 73.9 at 2,577 tokens mem9: 73.6 at 1,597 tokens Supermemory: 73.5 at 15,238 tokens MemoryLake: 72.5 at 5,202 tokens Viking: 69.3 at 5,964 tokens Zep / Graphiti: 63.8 at 1,862 tokens MemOS 88.8Mem0 77.7 Memoria: 79.0 at 1,369 tokens Memoria 79.0
LongMemEval-S Memoria sends 8,085 tokens per question; 5 of 13 send more
50 60 70 80 90 1k 10k 100k context tokens per question (log scale) → ↑ accuracy (%) MemOS: 89.2 at 4,151 tokens LCM-X: 87.0 at 5,164 tokens EverOS: 80.4 at 12,379 tokens Zep / Graphiti: 79.8 at 117,106 tokens mem9: 78.0 at 3,805 tokens Letta: 77.7 at 49,431 tokens Hindsight: 72.2 at 29,755 tokens Supermemory: 66.1 at 6,635 tokens MemMachine: 63.6 at 2,803 tokens Viking: 61.1 at 2,291 tokens Mem0: 56.0 at 856 tokens Cognee: 51.8 at 10,305 tokens MemOS 89.2Mem0 56.0 Memoria: 81.4 at 8,085 tokens Memoria 81.4
BEAM-100K Memoria sends 19,084 tokens per question; 3 of 12 send more
50 60 70 80 90 1k 10k 100k context tokens per question (log scale) → ↑ nugget score (%) LCM-X: 71.8 at 11,926 tokens Viking: 70.8 at 2,023 tokens Mem0: 70.4 at 1,055 tokens Hindsight: 70.2 at 24,085 tokens Letta: 69.2 at 82,013 tokens Zep / Graphiti: 68.7 at 8,667 tokens MemOS: 66.9 at 1,636 tokens Supermemory: 66.0 at 6,294 tokens mem9: 65.8 at 5,372 tokens MemMachine: 64.8 at 5,413 tokens Cognee: 59.3 at 33,065 tokens EverOS: 58.6 at 7,393 tokens LCM-X 71.8Mem0 70.4 Memoria: 71.4 at 19,084 tokens Memoria 71.4

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 were
Under 1 day: 15.8% of used memories 16% Under 1 day 1–7 days: 26.2% of used memories 26% 1–7 days 7–30 days: 26.2% of used memories 26% 7–30 days 30–90 days: 15.8% of used memories 16% 30–90 days Over 90 days: 15.8% of used memories 16% Over 90 days
Where they came from
Notes, documents, email: 43% of used memories Notes, documents, email 43% Older agent memory: 26.7% of used memories Older agent memory 27% Earlier session, same project: 16.3% of used memories Earlier session, same project 16% Earlier session, other project: 7.7% of used memories Earlier session, other project 8% Same session: 6.3% of used memories Same session 6%

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 Memoria With 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.
Full method, results and limits →