LoCoMo
1,986 questions · 272 dialogue-session documents
R@1
67.1%
R@5
90.1%
R@10
94.7%
MRR
0.768
Median
<1 ms
P95
<1 ms
Strongest result. Temporal and open-domain categories remain lower than direct recall.
Public benchmark · rerun July 23, 2026
On the benchmark-native LongMemEval retrieval task, hippocampOS finds labelled evidence in its first five results 97.0% of the time—edging MemPalace Raw without an LLM, provider, or network call.
Headline
97.0%
LongMemEval scoped Recall@5
Providers
None
Raw edge
+0.4 pp
12,865
public questions
3
benchmark families
10
retrieved results
0
provider or network calls
Latest results
Recall@K asks whether any labelled evidence appears in the first K results. MRR rewards placing the first correct result nearer the top.
1,986 questions · 272 dialogue-session documents
R@1
67.1%
R@5
90.1%
R@10
94.7%
MRR
0.768
Median
<1 ms
P95
<1 ms
Strongest result. Temporal and open-domain categories remain lower than direct recall.
500 questions · benchmark-native evidence-session scope
R@1
—
R@5
97.0%
R@10
98.4%
MRR
0.934
Median
Not reported
P95
Not reported
The same full 500-question scoped run used for the 97.0% headline and the direct MemPalace Raw comparison.
10,379 questions · 12,394 dialogue-evidence documents
R@1
39.9%
R@5
58.4%
R@10
65.6%
MRR
0.479
Median
96 ms
P95
129 ms
P95 latency is 129 ms. Assistant-fact evidence reaches 86.7% R@10; implicit connections remain weakest at 28.2%.
Latest measured values from the public benchmark run on July 23, 2026.
| Benchmark | R@1 | R@5 | R@10 | MRR | Median | P95 |
|---|---|---|---|---|---|---|
| LoCoMo | 67.07% | 90.08% | 94.71% | 0.7676 | <1 ms | <1 ms |
| LongMemEval-S (scoped) | Not reported | 97.00% | 98.40% | 0.9340 | Not reported | Not reported |
| ConvoMem sample | 39.89% | 58.36% | 65.62% | 0.4790 | 96 ms | 129 ms |
This separate 19,195-session run removes the question-supplied scope. Direct assistant facts and explicit updates retrieve well; preferences, cross-session synthesis, and temporal evidence remain weak under global interference. These category values do not power the 97.0% scoped headline.
Same LongMemEval-S dataset. Same 500 questions. Same evidence-session recall-any@5 metric. Both runs use each question’s supplied retrieval scope and require no LLM. Under those aligned conditions, the comparison is meaningful.
hippocampOS
97.0%
R@5 · full 500 questions
hippocampOS depth
98.4%
R@10
0.934
MRR
0
provider calls
Important boundary
MemPalace also reports 98.4% R@5 for its Hybrid v4 system on a held-out 450-question split. That is a stronger published number under a different experimental condition; it is not the Raw baseline compared above.
Harder stress test
50.0% R@5 across all 19,195 sessions.
The global run removes each question’s supplied haystack and searches the entire parsed corpus. It tests interference at scale, not the benchmark-native condition used in the direct comparison.
hippocampOS and MemPalace Raw use the same full 500-question retrieval protocol. MemPalace Hybrid v4 uses a held-out split; the remaining systems publish generated-answer accuracy or use different datasets.
| System | Published result | Benchmark | Metric | Conditions |
|---|---|---|---|---|
| hippocampOS | 97.0% R@5 | LongMemEval-S | Evidence-session recall-any@5 | Full 500 questions · benchmark-native scope · no LLM |
| MemPalace Raw | 96.6% R@5 | LongMemEval-S | Evidence-session recall-any@5 | Full 500 questions · benchmark-native scope · no LLM |
| MemPalace | 98.4% R@5 | LongMemEval-S | Evidence-session recall@5 | Hybrid v4 · held-out 450-question split |
| Mem0 v3 | 94.8% overall | LongMemEval | Generated-answer evaluation at top 50 | Managed memory · LLM answerer and judge |
| Zep / Graphiti | 94.8% | Deep Memory Retrieval | Answer accuracy | Different dataset and evaluation pipeline |
Methodology