Training-Free Lexical–Dense Fusion for Conversational-Memory Retrieval

Reproduction code and paper for Training-Free Lexical–Dense Fusion for Conversational-Memory Retrieval (Christian Lysenstøen, 2026).

TL;DR. On the LoCoMo benchmark, fusing BM25 with turn-level late-interaction dense retrieval (max-similarity over per-turn vectors) at the score level — with a single leave-one-conversation-out weight — beats either component, needs no training, and runs on CPU. Adding a cross-encoder reranker on top hurts. We credit the turn-level late-interaction retriever to Nano-Memory (Wu et al., 2026); the contribution here is a controlled study of the lexical–dense fusion recipe around it.

Headline result

LoCoMo, Hit@1, honest leave-one-conversation-out cross-validation (e5-large-v2):

Method	Hit@1
BM25	0.640
Dense, mean-pool session (early interaction)	0.427
Dense, max-sim turn (late interaction)	0.664
BM25 ⊕ max-sim turn (fusion)	0.752

• Fusion adds +8.8 to +17.2 pp Hit@1 over late interaction alone across six encoders (all p < 1e-4).
• An off-the-shelf web-search cross-encoder reranker on the fused list degrades Hit@1 by 6.9 pp.
• Honest boundary: on LongMemEval-S — a lexical regime where BM25 saturates — the net fusion gain over BM25 is small and not significant.

Holding the retrieval unit fixed at session and reusing identical cached turn vectors, switching the pooling operator from early (mean-pool) to late (max-sim) lifts dense Hit@1 above the BM25 reference at every encoder.

A division of labor: dense late interaction wins on multi-hop and temporal questions, trails BM25 on adversarial ones, and fusion combines whichever signal a query needs.

Full paper: paper/paper.pdf. arXiv: link to be added.

Repository structure

paper/            LaTeX source, BibTeX, .bbl, figures, and the compiled PDF
results/          result receipts (JSON + Markdown) and figures for every table
data/locomo/      where to place the LoCoMo benchmark (see its README)
*.py              reproduction scripts (see "Reproduction" below)
requirements.txt

Reproduction

pip install -r requirements.txt
# 1. Get LoCoMo  ->  see data/locomo/README.md
# 2. Build cached turn/query embeddings (first run only):
python embed_locomo.py && python embed_turns.py
# 3. Run the experiments (CPU, from cache):
python tune13_interaction.py        # late vs early + fusion        (Tables 1,2,4,5,7)
python tune13b_fusion_vs_late.py    # fusion vs late-alone          (Table 3)
python tune10_rerank.py             # cross-encoder reranking       (Table 6)
python analysis_deep.py             # per-category, length, alpha   (Tables 9,10; Figs 2,3)
python lme_interaction.py --model intfloat/e5-base-v2 \
       --qpref "query: " --ppref "passage: "   # LongMemEval-S      (Table 8)
python make_figures.py              # Figure 1

Every number in the paper has a receipt under results/<run>/ (*.json raw, *.md human-readable). Embedding caches are regenerated by the embed_* scripts and are not checked in.

Key modules

File	Purpose
locomo.py	LoCoMo loader (LongMemEval-S loads via the HF datasets Hub)
embed_locomo.py, embed_turns.py, embed_turns_multi.py	build & cache turn / query embeddings
tune7_bm25.py … tune13_interaction.py, tune13b_fusion_vs_late.py	LoCoMo experiments (BM25, fusion, ablations); BM25 + metrics are defined inline
tune10_rerank.py	cross-encoder reranking (the negative result)
lme_interaction.py, lme_maxsim.py	LongMemEval-S
analysis_deep.py, latency.py, make_figures.py	analysis + figures
e2e_qa.py	optional end-to-end QA harness (needs an LLM reader API key)

Citation

@misc{lysenstoen2026fusion,
  title  = {Training-Free Lexical--Dense Fusion for Conversational-Memory Retrieval},
  author = {Lysenst{\o}en, Christian},
  year   = {2026},
  note   = {arXiv preprint}
}

License

Code is released under the MIT License (see LICENSE). The paper in paper/ is licensed CC BY 4.0.