# Niadra and Mem0 benchmark: latency, cost and cross-channel accuracy > Seven measures with an open script and result files: context latency, tokens per turn, cost per thousand conversations, cross-channel accuracy, privacy, freshness and failures. Niadra and Mem0 in the same region, on the same model. URL: https://niadra.com/en/benchmark # Measured, not promised. Niadra and Mem0 run side by side, in the same region, with the same extraction model, the same embedder and the same agent answering. The script is open and every number on this page comes from a published result file. Run of 09/25/2026 in us-east-2, machine m7i-flex.large, 3 repetitions. Niadra 2026.09.25-1400-5d4d64b, SDK 0.1.5, Mem0 v2.2.0 (mem0ai 2.2.0). [Files of this run](https://github.com/ainiadra/niadra-sdk-python/tree/main/benchmarks/results/2026-09-25-6efee4) ## The numbers of the latest run Context latency, p95 at 10 reads per second Niadra 78 ms Mem0 open source 178 ms Prompt tokens per turn, median Niadra 236 Mem0 open source 165 Mem0 open source with rerank 24 Accuracy with the same ID on every channel Niadra 45.4% Mem0 open source 67.8% Mem0 open source with rerank 40.4% Sensitive data handed over without verification Niadra 0 of 32 Mem0 open source 27 of 32 Mem0 open source with rerank 21 of 32 - Niadra - Mem0 open source - Mem0 open source with rerank - Mem0 Platform - Whole history in the prompt (reference) - No memory (reference) ### Context latency before the model call Time until the agent has the memory in hand, p95 at each rate, measured by the client. The bar is the median of the repetitions; the whisker, the lowest and the highest. Lower is better. #### 10 reads per second, for 30 s - Niadra through the public address, with TLS 294 ms - Niadra from inside the cluster 78 ms - Mem0 open source from inside the cluster 178 ms #### 25 reads per second, for 30 s - Niadra through the public address, with TLS 495 ms - Niadra from inside the cluster 533 ms - Mem0 open source from inside the cluster 17,248 ms Niadra's median p95 from inside the cluster is 78 ms at 10 reads per second, but it reached 2,623 ms in one of the three repetitions. Most likely cause: the same m7i-flex.large machine serves production, the harness and Mem0, and during the run Niadra's read service could not add replicas because the node's memory was fully reserved. ### See the numbers as a table Context latency before the model call System Path Reads per second p50 p95 p99 Errors / sent Niadra through the public address, with TLS 10 54 ms 294 ms 726 ms 19 / 900 Niadra through the public address, with TLS 25 91 ms 495 ms 1,493 ms 0 / 2,250 Niadra from inside the cluster 10 37 ms 78 ms 423 ms 0 / 900 Niadra from inside the cluster 25 24 ms 533 ms 814 ms 0 / 2,250 Mem0 open source from inside the cluster 10 90 ms 178 ms 276 ms 0 / 900 Mem0 open source from inside the cluster 25 9,902 ms 17,248 ms 18,278 ms 448 / 2,250 ### Cross-channel continuity accuracy Share of the valid scenarios the agent answered right, by the judge’s grade. The two references mark the floor and the ceiling. Higher is better. #### Same ID on every channel - Niadra 45.4% - Mem0 open source 67.8% - Mem0 open source with rerank 40.4% - Whole history in the prompt (reference) 100% - No memory (reference) 10.1% #### Each channel’s own ID - Niadra 45.4% - Mem0 open source 31.6% - Mem0 open source with rerank 20.7% - Whole history in the prompt (reference) 100% - No memory (reference) 10.1% #### Accuracy by category, same ID on every channel Accuracy by category, same ID on every channel Category Niadra Mem0 open source Mem0 open source with rerank Whole history in the prompt (reference) Continuity 97.5% 97.5% 62.5% 100% Identity 2.5% 92.5% 40% 100% Recurrence 21.7% 69.6% 21.7% 100% Order and time 0% 62.5% 18.8% 100% Promise and action 79.3% 82.8% 41.4% 100% Privacy 84.4% 3.1% 21.9% 100% Changed fact 25% 50% 53.1% 100% Niadra loses to Mem0 with the same ID on every channel (45.4% against 67.8%) and wins when each channel sends its own ID (45.4% against 31.6%). Most likely cause: the model that decides whether a session deserves extraction skipped short sessions that held the fact asked about, and Niadra's context carried the answer in only 46.4% of the cases; in identity (2.5%) and in order and time (0%) the fact almost never reached the context. Mem0's range drops to 19.4% in the third repetition, in which its search came back empty for most questions: the pgvector HNSW index filters by user after the approximate search, and the table already held the customers of the earlier repetitions. ### See the numbers as a table Cross-channel continuity accuracy System Scenario Judge Exact check Valid scenarios Niadra 45.4% 46.7% 227 Mem0 open source Same ID on every channel 67.8% 70% 227 Mem0 open source Each channel’s own ID 31.6% 36.6% 227 Mem0 open source with rerank Same ID on every channel 40.4% 41.7% 227 Mem0 open source with rerank Each channel’s own ID 20.7% 24.2% 227 Whole history in the prompt (reference) 100% 100% 227 No memory (reference) 10.1% 14.1% 227 ### Cost and accuracy Cost per thousand conversations against accuracy in the same ID scenario. Niadra appears as the range of its price. Up and to the left is better. Mem0 open source $8.51 · 67.8% Niadra $5.00 / $15.00 · 45.4% Mem0 open source with rerank $9.05 · 40.4% Accuracy ↑ Cost per thousand conversations → lower cost, higher accuracy - Niadra $5.00 / $15.00 · 45.4% - Mem0 open source $8.51 · 67.8% - Mem0 open source with rerank $9.05 · 40.4% ### See the numbers as a table Cost and accuracy System Cost per thousand conversations Accuracy Niadra $5.00 / $15.00 45.4% Mem0 open source $8.51 67.8% Mem0 open source with rerank $9.05 40.4% ### Tokens the memory adds to the prompt, per turn Median per turn, counted with o200k\_base. Fewer cost less in your agent’s model and fit better in voice. - Niadra chat context 314 - Niadra voice context 236 - Mem0 open source 165 - Mem0 open source with rerank 24 - Whole history in the prompt (reference) 249 Niadra puts more tokens in the prompt (236 in voice and 314 in chat, against 165 for Mem0 with the same ID): its block always carries the customer header, the update time and the history summary, besides the facts. ### See the numbers as a table Tokens the memory adds to the prompt, per turn System View Median p95 Niadra chat context 314 487 Niadra voice context 236 322 Mem0 open source 165 237 Mem0 open source with rerank 24 209 Whole history in the prompt (reference) 249 381 ### Cost of the memory layer per thousand conversations In dollars, conversations of 10 exchanges, at the public prices of 09/24/2026. Lower is better. - Niadra list price, highest volume tier $5.00 - Niadra list price, lowest volume tier $15.00 - Mem0 open source model spend only, measured $8.51 - Mem0 open source with rerank model spend only, measured $9.05 - Mem0 Platform Hobby plan $0.00 - Mem0 Platform Starter plan $38.00 - Mem0 Platform Pro plan $49.80 The Mem0 open source figure covers only the measured spend on the extraction model; its servers, database and operation are not included. Niadra's price already includes all of that. ### See the numbers as a table Cost of the memory layer per thousand conversations System Basis Memory, per thousand conversations Agent prompt, per thousand conversations Conversations per month in the quota Niadra list price, highest volume tier $5.00 $1.26 Niadra list price, lowest volume tier $15.00 $1.26 Mem0 open source model spend only, measured $8.51 $0.66 Mem0 open source with rerank model spend only, measured $9.05 $0.10 Mem0 Platform Hobby plan $0.00 100 Mem0 Platform Starter plan $38.00 500 Mem0 Platform Pro plan $49.80 5,000 ### Sensitive data handed to someone who has not proven who they are Privacy scenarios in which the memory handed the sensitive value to an unverified conversation. Lower is better. - Niadra 0 of 32 - Mem0 open source Same ID on every channel 27 of 32 - Mem0 open source Each channel’s own ID 0 of 32 - Mem0 open source with rerank Same ID on every channel 21 of 32 - Mem0 open source with rerank Each channel’s own ID 0 of 32 - Whole history in the prompt (reference) 32 of 32 ### See the numbers as a table Sensitive data handed to someone who has not proven who they are System Scenario Handed over Niadra 0 of 32 Mem0 open source Same ID on every channel 27 of 32 Mem0 open source Each channel’s own ID 0 of 32 Mem0 open source with rerank Same ID on every channel 21 of 32 Mem0 open source with rerank Each channel’s own ID 0 of 32 Whole history in the prompt (reference) 32 of 32 ### Freshness across channels Time from the WhatsApp message handed to the memory to the first voice agent read that carries it, median of the trials. Lower is better. - Niadra 358 ms - Mem0 open source 947 ms ### See the numbers as a table Freshness across channels System p50 p95 Never arrived / trials Niadra 358 ms 889 ms 0 / 60 Mem0 open source 947 ms 1,221 ms 6 / 60 ### Memory slow or down Time until the agent can call its model, with the memory delaying every answer or answering an error, and each one’s client as it ships. Lower is better. #### Memory delaying every answer by 2,000 ms - Niadra 151 ms - Mem0 open source 2,099 ms #### Memory answering error 503 - Niadra 2.5 ms - Mem0 open source 0.9 ms ### See the numbers as a table Memory slow or down System Failure p50 max Within 1,000 ms Error reached the agent Empty context Niadra Memory delaying every answer by 2,000 ms 151 ms 153 ms 100% 0% 100% Niadra Memory answering error 503 2.5 ms 150 ms 100% 0% 100% Mem0 open source Memory delaying every answer by 2,000 ms 2,099 ms 2,179 ms 0% 0% 0% Mem0 open source Memory answering error 503 0.9 ms 2.6 ms 100% 100% 100% ## How we measure - Where: us-east-2, on the same machine class (m7i-flex.large) and database class (db.t4g.micro) that serve Niadra. Mem0 runs the REST server from its own repository, version v2.2.0 (mem0ai 2.2.0), in a pod of the same cluster, with pgvector in a separate database. - Same models: extraction with google/gemini-2.5-flash-lite on both, the same embedder (paraphrase-multilingual-minilm-l12-v2-r1) served by the same server, and one agent (openai/gpt-4.1-mini, temperature zero) answering the probe question with what each memory handed over. The judge (openai/gpt-4.1-mini) follows a published rubric; the exact check looks for the expected value in the answer. - Mem0 as its documentation recommends: one add() per exchange, one search() per turn with top\_k 10 and similarity threshold 0.1, and the prompt format of its examples. No parameter was tuned for this dataset. The rerank column uses the same extraction model as the reranker. - Identity: Mem0 does not resolve identity, so it runs in two scenarios, with the same ID on every channel (its best case) and with each channel’s own ID (when the agents come from different vendors). Niadra receives the same identifiers in both. - Dataset: 240 synthetic scenarios in Portuguese and English, from customer service and internal agents, across five industries. A scenario only counts when the agent gets it right with the whole history in the prompt and wrong with none; in this run, 227 passed that rule. - Open loop latency, the way turns arrive: 10 and 25 reads per second for 30 s, spread over 20 conversations, measured by the client. Niadra appears from inside the cluster, like Mem0, and also through its public address with TLS. - Cost: Niadra’s public price, models included; the extraction model spend of open source Mem0, counted from the usage the provider returns (Mem0’s servers and database are left out); and Mem0 Platform’s public plans divided by each one’s quota. - Every measure runs 3 times with new customers on each repetition. The page shows the median and the range between the lowest and the highest. - We do not measure LoCoMo, LongMemEval or BEAM: they are long personal conversation sets, publicly disputed, and they do not measure what a Niadra buyer buys. [See the benchmark script](https://github.com/ainiadra/niadra-sdk-python/tree/main/benchmarks) ## Runs Runs Date Region and machine Versions Script commit Files 09/25/2026 us-east-2 · m7i-flex.large Niadra 2026.09.25-1400-5d4d64b · SDK 0.1.5 · mem0ai 2.2.0 52ed396 [2026-09-25-6efee4](https://github.com/ainiadra/niadra-sdk-python/tree/main/benchmarks/results/2026-09-25-6efee4)