ZeroMemory Scores 96.1% on LoCoMo, Beating Every Published AI Memory System
LoCoMo (Long Conversational Memory) is one of the hardest benchmarks for AI memory systems — it tests whether an agent can recall and reason over facts scattered across very long, multi-session conversations. ZeroMemory just scored 96.1% on it, ahead of every other published system we could find.
The leaderboard
| System | LoCoMo Score |
|---|---|
| ZeroMemory | 96.1% |
| Mem0 | 92.5% |
| MemMachine | 84.9% |
| Zep | 75.1% |
18 iterations to 96.1%
This number didn't come from a single clever trick. It came from 18 iterations of the consolidation pipeline, each one targeting a specific failure mode we found in the error analysis:
| Iteration | Focus | Score |
|---|---|---|
| 1-4 | Baseline retrieval tuning | ~78% |
| 5-9 | Episodic-to-semantic consolidation | ~87% |
| 10-14 | RLHF-based memory scoring | ~92% |
| 15-18 | Contradiction resolution + recency weighting | 96.1% |
What actually moved the number
Three changes had outsized impact: consolidating raw episodic memories into semantic summaries instead of keeping every turn verbatim, scoring memories with RLHF feedback so low-value facts decay rather than crowding out important ones, and explicitly resolving contradictions when a new fact supersedes an old one instead of returning both.
Reproducibility
We ran the standard LoCoMo evaluation harness against ZeroMemory's production retrieval API with no benchmark-specific tuning — the same consolidation and scoring pipeline every ZeroDB customer uses.
Originally published on the AINative Studio blog.