Loading your app…

LoCoMo benchmark

Ahead of Zep, LangMem, and Mem0 on LoCoMo.

Make0 Max and Standard versus Zep, LangMem, Mem0, and OpenAI. Full Context is omitted. SWE-ContextBench retrieval is on its own page.

Make0 Max mode · Experimental*

0.0%

Overall accuracy · 8s+ latency

Evaluated 10 Sep 2026

*100-question subset

Make0 Standard

0.0%

Overall accuracy · 1–3s latency

Evaluated 25 Oct 2026

Retrieval

0.0%

Hit@10 · Make0 Standard

LLM-as-a-judge

How Make0 compares on LoCoMo

Max mode is an experimental Make0 configuration with 8s+ retrieval latency and limited availability. Standard is available to all users and takes 1–3s. We are working to lower Max mode's latency. Competitor scores are published; Full Context is omitted. *Max mode was evaluated on a 100-question subset of LoCoMo, not the full 1,540 questions, so it is not directly comparable with full-benchmark scores.

Overall

Overall LoCoMo accuracy

LLM-as-a-judge score across all categories. Higher bars are better.

LoCoMo overall accuracy for Make0 Max, Make0 Standard, Zep, LangMem, Mem0, and OpenAI

Breakdown

Accuracy by category

Single-hop, multi-hop, world knowledge, temporal, and overall.

LoCoMo accuracy by system and reasoning category, excluding Full Context
MethodSingle-hopMulti-hopWorld knowledgeTemporalOverall
Make0 (Max mode · Experimental*)Evaluated 10 Sep 202694.4%90.6%76.9%97.3%92.0%
Make0 (Standard)Evaluated 25 Oct 202690.73%67.38%61.46%75.08%81.36%
Zep79.43%69.16%73.96%83.33%79.09%
LangMem74.47%61.06%67.71%86.92%78.05%
Mem062.41%57.32%44.79%66.47%62.47%
OpenAI63.79%42.92%62.29%21.71%52.9%

Make0 Standard retrieval: Hit@10 98.4%.

Best score in each column is highlighted.

Tinted Make0 cells beat every published baseline, even when Max mode* scores higher.

Retrieval cost

A fraction of the full conversation

Mean tokens per retrieval call on LoCoMo. Make0 Standard sends 2,629 memory tokens to the answering model, 90% fewer than passing the full conversation (26,031), and less than Zep (3,911) or OpenAI (4,437). Mem0's 2025 version retrieves less (1,764) and scores 62.47% overall, against 81.36% for Make0 Standard. Mem0's 2026 algorithm reports 6,956 tokens per call.

Retrieval cost

Mean tokens per retrieval

Context tokens sent to the answering model per LoCoMo question. Lower bars are better.

Mean retrieval tokens for Make0 Standard, Mem0 (2025 and 2026), Zep, OpenAI, and full-context

Make0 is the mean context tokens from the full Standard LoCoMo evaluation (evaluated 25 Oct 2026). Mem0 (2025), Zep, OpenAI, and Full-context figures are from Table 2 of Mem0's 2025 paper. Mem0 (2026) is self-reported on Mem0's research page.

Performance by category

Make0's clearest edge is single-hop recall.

These are lookup questions: a name, a plan, a reason, already said once in conversation.

  • What did the charity race raise awareness for?
  • What type of people does that adoption agency support?
  • What does Melanie think about Caroline adopting?

Standard leads the published baselines here at 90.73% versus Zep (79.43%). Max mode* leads every category on its 100-question subset. Standard is just behind Zep on multi-hop and trails Zep and LangMem on open-domain (“would she…?”) and temporal (“when did…?”) questions.

Single-hop reasoning (94.4%)

  • What did the charity race raise awareness for?
  • Why did Caroline choose that adoption agency?
  • What are her plans for the summer?

A fact stated in one sitting. Max mode* leads at 94.4%, ahead of Zep (79.43%) and LangMem (74.47%). Standard also leads the published baselines here at 90.73%.

Multi-hop reasoning (90.6%)

  • What activities does Melanie partake in?
  • Where has she camped?
  • What career path has Caroline decided to pursue?

Piecing together details that never appear in one session. Max mode* leads at 90.6%, above Zep (69.16%) and LangMem (61.06%). Standard scores 67.38%, ahead of LangMem but just behind Zep.

World knowledge / open-domain reasoning (76.9%)

  • Would Caroline likely have Dr. Seuss books on her bookshelf?
  • Would Melanie prefer a national park or a theme park?
  • What would Caroline’s political leaning likely be?

A stored fact plus an inference the transcript never states. Max mode* scores 76.9%, just above Zep (73.96%). Standard is weaker here at 61.46%, behind Zep and LangMem (67.71%).

Temporal reasoning (97.3%)

  • When did Caroline go to the LGBTQ support group?
  • When is Melanie planning on going camping?
  • How long has Caroline had her current group of friends?

When something happened, or how long ago. Max mode* leads at 97.3%, ahead of LangMem (86.92%) and Zep (83.33%). Standard scores 75.08%, behind those two published systems.

Methodology

What we measured, and what we left out.

Evaluation

LoCoMo, judged consistently

The benchmark covers single-hop, multi-hop, world knowledge, and temporal questions across long, multi-session conversations. Make0 answers and scoring both use GPT-5.6 Luna.

Standard · available now

The default setting for all users: 81.36% overall with 1–3s retrieval latency. Evaluated 25 Oct 2026.

Max · experimental*

A higher-capacity setting at 92.0% overall. It has 8s+ latency and limited availability while we work to reduce it. Evaluated 10 Sep 2026 on a 100-question subset of LoCoMo.

Full Context is excluded: passing an entire conversation into the prompt is not a retrieval-based memory system.

Notes

  1. *Max mode was evaluated on a 100-question subset of LoCoMo, not the full 1,540 questions, so it is not directly comparable with full-benchmark scores.
  2. Latency is how long it takes to retrieve memory from your Brain and return it to the client. It does not include the time the downstream model spends generating an answer. Max mode is experimental, takes 8s+ per retrieval, and has limited availability. Standard is available to all users at 1–3s. We are working to lower Max mode's latency.
  3. Make0 Max mode and Make0 Standard are two configurations of the same product. Standard scores were produced on 25 Oct 2026; Max mode scores on 10 Sep 2026. Both use GPT-5.6 Luna as both answerer and judge.
  4. The LoCoMo judge prompt is equivalent to the one used for Mem0's original paper comparison: it is generous about equivalent topic matches and date phrasing, including equivalent calendar periods and small off-by-one day differences.
  5. Open-domain is LoCoMo world-knowledge (JSON category 3, 96 questions). Single-hop is category 4 (841), multi-hop is category 1 (282), temporal is category 2 (321). Adversarial is not reported.
  6. Zep, LangMem, and Mem0 scores are from the published Memori LoCoMo Benchmark Results. OpenAI scores are from Mem0's original paper. They were not re-run in this Make0 evaluation. Full Context is omitted because it is not a retrieval-based memory system.
  7. Mean tokens per retrieval call for Make0 Standard is the mean memory context from the full LoCoMo evaluation. Mem0 2025 (1,764), Zep (3,911), OpenAI (4,437), and Full-context (26,031) are from Table 2 of Mem0's original paper. Mem0 (2026, 6,956) is Mem0's self-reported figure for its 2026 token-efficient algorithm on Mem0's research page; its accuracy is not shown here.

Start with one Brain

Free to start

Build your Brain in two minutes

Sign up, add a few memories, and connect your first agent. That's the whole setup.

Make0Your BrainClaudeChatGPTCursorCursorGrok BotMake0Make0 Agent