Urmzd Mukhammadnaim urmzd
urmzd
/ llmem_bench_main.rs
Created
April 17, 2026 00:13
LongMemEval retrieval comparison: MemPalace's 96.6% is ChromaDB's add+query. mnemonist gets 96.0% with a from-scratch Rust HNSW (3 days earlier). One script (reproduce_benchmark.sh) clones everything, builds, and runs end-to-end. All flaws documented.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| //! LongMemEval retrieval benchmark — mnemonist (Apr 1, 2026, commit 3ca2c39) | |
| //! | |
| //! Uses the exact same methodology as MemPalace's raw-mode benchmark: | |
| //! | |
| //! 1. For each of 500 questions, build a per-question haystack (~48 sessions) | |
| //! 2. Index only user turns, concatenated per session | |
| //! 3. Embed haystack + query (fastembed/ONNX, all-MiniLM-L6-v2) | |
| //! 4. Retrieve top-5 by cosine similarity via native HNSW index | |
| //! 5. Score: is the gold session in the top-5? | |
| //! |