/ ml (SlashML) Benchmarks: Accurate and really fast.
/ ml is a native memory engine with graph, temporal, and code reasoning, evaluated head-to-head across LoCoMo, 600-task DolphinBench, Official LongMemEval, and 10M-token ICLR 2026 BEAM.
0.86 ms at 26K LoCoMo; 41.7 ms at 10M BEAM). Add ~10-25 ms HTTPS RTT on Managed Cloud or <2 ms intra-VPC on / ml sovereign.
LoCoMo · 94.5% in 0.86ms
DolphinBench · 90.5% in 33.9s
BEAM Scale · +35.7 pts & 240×+ faster @ 10M
Read full benchmark methodology, academic citations & reproduction notes
- * Mem0 (arXiv:2504.19413): Top-200 Retrieval Configuration & 10M Scale Profile: In Mem0’s published paper, standard
Top-10 Retrievalscores 66.9% (68.4%with Mem0g Graph Memory on the $249/mo Pro tier). Their blog’s92.5%benchmark usesTop-200 Retrieval(~6,956 tokens/queryinjected per prompt, increasing downstream LLM token usage ~5× and 26K-token P50 latency to 3,550 ms). On 10M-token BEAM (12.3M tokens,20,561 chunks), Mem0 scores 48.6% with recall latency exceeding >10,000 ms and 16–40+ hours of LLM ingestion calls (vs / ml: 84.3% in 41.7 ms, 3.7× fewer tokens/query, and 2.4 min synchronous ingestion). - * Honcho (Plastic Labs): Query-Time Dialectic LLM Synthesis & Background Workers: Honcho’s
89.0%LoCoMo score uses a query-time/chatDialectic LLM reasoning endpoint alongside background “dreaming” LLM synthesis workers, adding 1–5s of LLM inference latency (~2,500 ms P50) and secondary token billing per recall. On 600-task DolphinBench, Honcho scores 68.5% at $142.85 (3.1× higher run cost than / ml). - * Hindsight (Vectorize, arXiv:2512.12818): Backbone Model Configuration & Tail Latency: Hindsight’s paper reports 83.6% (standard) / 85.6% (large) and reaches
89.6%when pairing multi-pass agenticreflectloops with a frontier answering backbone. On DolphinBench it scores 69.5% ($84.65) with 158.9s–602.9s (10 min) P95 tail latency. - * Zep / Graphiti (arXiv:2501.13956): Async Graph Ingestion & 10M Traversal Profile: Zep scored 58.4% in Mem0’s evaluation (reflecting async graph ingestion queue delay when queried immediately) and 75.1% in Zep’s 4-category re-run, with 1,290 ms P50 graph traversal latency at 26K tokens. On 10M-token BEAM (ICLR 2026), Zep scores 42.1% with multi-hop graph traversal exceeding >8,000 ms per query and requiring 16–40+ hours of LLM entity/edge extraction (vs / ml: 84.3% in 41.7 ms and 2.4 min synchronous ingest).
- * Supermemory: DolphinBench Multi-Session Evaluation Profile: Supermemory self-reports
81.6%–85.2%on LoCoMo/LongMemEval subsets. On the independent 600-task DolphinBench coding & tool-use suite, Supermemory scores 59.2% (compared to 65.7% for the Built-in no-memory baseline), with a total benchmark run cost of $345.64 (5.6× baseline cost).
Audited accuracy, latency, and token economics.
Every metric below comes from the published configurations and audited runs cited above, across short-horizon recall, 600-task agent execution, 10M-token scaling, and surgical GDPR deletion.
| Metric / Suite | / ml (SlashML) | Wrapper / Graph Stacks | Key Advantage |
|---|---|---|---|
| LoCoMo Retrieval (ACL 2024)#1 SOTA | 94.5% (Top-20) · 0.86 ms P50 | Mem0 Top-200: 92.5% (3,550 ms) · Honcho /chat: 89.0% (2,500 ms) · Hindsight: 85.6% (1,500 ms) · Zep: 75.1% (1,290 ms) | 4,128× faster than Mem0 Top-200 |
| DolphinBench (600 multi-session tasks) | 90.5% · 33.9 s P50 ($45.90) | Mem0: 70.7% (37.7 s) · Hindsight: 69.5% (55.3 s, $84.65) · Honcho: 68.5% (44.7 s, $142.85) · Supermemory: 59.2% (52.7 s, $345.64) | +19.8 pts vs Mem0 · 3.1× cheaper than Honcho |
| Official LongMemEval | 90.0% | Full multi-session temporal & cross-turn evaluation | Consistent accuracy without prompt bloat |
| BEAM Scale (ICLR 2026 · 1M to 10M tokens) | 88.4% @ 1M (10.9 ms) · 84.3% @ 10M (41.7 ms) | Mem0 10M: 48.6% (>10,000 ms) · Zep 10M: 42.1% (>8,000 ms) | +35.7 pts & 240×+ faster @ 10M |
| 10M-Token Ingestion (12.3M tokens / 20,561 chunks) | 2.4 min (synchronous ingest) | 16–40+ hrs (async entity/edge extraction queue) | 400×–1,000× faster bulk indexing |
| Downstream Prompt Tokens / Query | ~1,400 tokens/query | ~6,956 tokens/query (Mem0 Top-200) | 80% lower downstream LLM bill (5× fewer tokens) |
| Surgical Cryptographic GDPR Unlearning | 6.41 ms | Full-index rebuild or async queue compaction | Instant verifiable erasure per user/fact |
All / ml tiers (including Free and $19/mo Pro) include native Graph & Relational Memory and zero-lag synchronous indexing. Read the full 12-benchmark technical report for verbatim traces and methodology.
Give your agents memory that keeps up.
Run the exact same native memory engine on Cloud or inside a single stateless Private VPC binary. Connect Claude Code, Cursor, LangGraph, or your custom stack in under a minute.