memweave — R@5 98.00%, 100% recall at R@23 | local embeddings (all-MiniLM-L6-v2), no LLM | full reproduction setup
#2
by sachinsharma9780 - opened
Hi!
I have evaluated memweave (https://github.com/sachinsharma9780/memweave) on
LongMemEval-S and wanted to share results + full reproduction setup in case
it's useful for others benchmarking retrieval systems.
Results — held-out split (450 questions)
| K | Recall@K | NDCG@K |
|---|---|---|
| 1 | 90.00% | 90.00% |
| 3 | 96.44% | 93.45% |
| 5 | 98.00% | 93.75% |
| 10 | 99.11% | 93.76% |
| 23 | 100.00% | 93.83% |
5-seed stratified cross-validated mean: R@5 = 97.24% ±0.12% — confirming stability
across different splits.
Setup
- Embedding model:
all-MiniLM-L6-v2via Ollama (local) - No LLM, no API key, no cloud at any stage
- Only user turns indexed (assistant turns excluded)
How
Three lightweight post-processors built on top of memweave's plugin API
(register_postprocessor()), with no changes to the core library:
- ECR — confidence-adaptive entity boost (additive)
- IDF — per-question corpus-relative keyword boost (multiplicative)
- CAATB — temporal proximity boost for time-offset queries (additive)
Parameters tuned on a 50-question dev set only. Held-out evaluated once —
no post-hoc adjustments.
Reproduction
Full setup, strategy source, and step-by-step commands:
https://github.com/sachinsharma9780/memweave/tree/main/benchmarks
Happy to answer questions about the pipeline or methodology.
sachinsharma9780 changed discussion title from memweave — R@5 98.00%, 100% recall at R@23 | local embeddings (all-MiniLM-L6-v2), no LLM to memweave — R@5 98.00%, 100% recall at R@23 | local embeddings | full reproduction setup (all-MiniLM-L6-v2), no LLM
sachinsharma9780 changed discussion title from memweave — R@5 98.00%, 100% recall at R@23 | local embeddings | full reproduction setup (all-MiniLM-L6-v2), no LLM to memweave — R@5 98.00%, 100% recall at R@23 | local embeddings (all-MiniLM-L6-v2), no LLM | full reproduction setup