Most agent memory systems use cosine similarity and hope for the best. SuperLocalMemory is the only one with peer-reviewed mathematical foundations, local-first architecture, and zero cloud dependency in the data path.
Every claim is sourced from public documentation. SLM claims carry arXiv citations.
| Feature | SLM V3.6 | Mem0 | Zep | LangMem | ChatGPT Memory |
|---|---|---|---|---|---|
| Mathematical retrieval | Fisher-Rao metric | — cosine | — cosine | — cosine | — proprietary |
| Local-first | ✓ fully local | ◑ Cloud + local | ✗ Cloud | ✗ Cloud | ✗ Cloud |
| No telemetry | ✓ | — | — | — | — |
| Open source | ✓ AGPL v3 | ✓ | ✗ | ✓ | ✗ |
| Peer-reviewed | ✓ 3 arXiv papers | — | — | — | — |
| Prompt compression | ✓ 60–95% | — | — | — | — |
| LLM cache | ✓ exact + semantic | — | ◑ partial | — | — |
| MCP protocol | ✓ 17+ tools | ✓ | ◑ partial | ✓ | — |
| Free tier | ✓ unlimited local | ◑ limited | ◑ limited | ◑ limited | ◑ limited |
| LoCoMo benchmark | 74.8% (local) / 87.7% (full) | not published | not published | not published | not published |
Competitor data sourced from public documentation as of June 2026. See arXiv:2603.14588 for SLM methodology. ◑ = partial support. If any data is incorrect, open an issue on GitHub.
Not features on a roadmap. Shipped, benchmarked, and cited in peer-reviewed literature.
Only SLM uses Fisher-Rao geodesics on a statistical manifold. Cosine similarity treats every memory as equally confident — it ignores the signal in your confidence scores entirely. Fisher-Rao does not.
Zero cloud in the data path. Mem0's local mode still routes through their SDK — which can send embeddings externally. SLM's local mode runs entirely on-device. There is no network call possible. Your memories are yours.
Three arXiv papers with reproducible benchmarks. LoCoMo 74.8% local-first — the highest published score without cloud dependency. Others publish blog posts with no methodology. We publish math.