Skip to content

Latest commit

 

History

History
38 lines (28 loc) · 1.85 KB

File metadata and controls

38 lines (28 loc) · 1.85 KB

LoCoMo Retrieval Bake-off — Matched ~2000-Token Budget

Date: 2026-07-18 22:27 UTC
Model (answer + judge): openai/gpt-4o-mini
Conversation: conv-26 (419 turns)
Questions: 15 (categories 1–4; category 5 skipped)
Shared context budget: ~2000 tokens (tiktoken cl100k_base)
Cogram promotion: novelty + surprise-mass (every content token; cap by Σ incident edge weight). Replaces the old frequency gate (count>=2).
Cogram decay: experienced ticks (1 tick per LoCoMo session, tick=19), not wall-clock days.

Per-category counts

Category Count
1 7
2 7
3 1

Results (sorted by accuracy)

Retriever Represents Accuracy Avg ctx tokens Build cost
cooc_ppr SPRIG / MemGraph-style PPR 40.0% (6/15) 1,996 0 (statistics)
bm25 Lexical / agent-knowledge 33.3% (5/15) 1,994 0 (statistics)
cogram_bm25 Cogram substrate + pluggable BM25 ranker 26.7% (4/15) 1,994 0 (statistics)
cogram Cogram (zero-LLM concept graph) 20.0% (3/15) 1,984 0 (statistics)
random Random floor 20.0% (3/15) 1,988 0 (statistics)
recency Naive recency 6.7% (1/15) 1,999 0 (statistics)

Headline

At equal ~2000-token budget, Cogram ranks #4 among zero-LLM retrievers (20.0% accuracy). Best zero-LLM retriever: cooc_ppr (40.0%).

Honest read

The count>=1 promotion fix raised Cogram from 16.7% (old benchmark_qa, count>=2) to 20.0% at matched budget. BM25 (33.3%) beats Cogram (20.0%) at equal budget — lexical matching is strong on LoCoMo factoid QA. Cogram is not the top zero-LLM retriever here; #4 of 6 at 20.0%. Vector/embedding baseline omitted for time; it is also the only baseline with a non-zero build cost.

Raw per-question data: benchmark_bakeoff_raw.json