← All systems
Zep (Graphiti)
OrganizationZep AI (YC W24)
Founded2023
FundingYC $500K verified; no later round found
CodeApache-2.0 · 29.2k★
Bi-temporal knowledge graph (Graphiti) with hybrid semantic + BM25 + graph retrieval.
Every number on file
| claimed | benchmark | config | runner | status |
|---|---|---|---|---|
| 71.2 / 63.8 | LongMemEval-S | gpt-4o / 4o-mini reader | self (paper) | SELF-REPORTED |
| 75.14 ± 0.17 | LoCoMo | corrected downward from ~84 after dispute | self (blog) | DISPUTED |
| 94.7 / 90.2 | LoCoMo / LME | gpt-5.4 reader + gpt-5.4 CoT judge · undated | self (research page) | SELF-REPORTED |
| 58.44 ± 0.20 | LoCoMo | 10 runs, Cat-5 excluded | Mem0 re-run | DISPUTED |
| DNF (>2 days) | LongMemEval | unified harness, Qwen2.5-7B | survey (arXiv:2604.01707) | THIRD-PARTY |
The only vendor that publicly corrected its own number (75.14). The 2025 dispute with Mem0 is the canonical case of config-dependent scoring.
Rows are not comparable — configs differ. A row graduates to a verdict only through a registered re-run. The protocol →