Cannot Reproduce LoCoMo Benchmark Results

[wangyu-ustc]

LoCoMo Benchmark Results - Significant Accuracy Gap

Issue: EverMemOS achieves 38.38% accuracy vs paper's claimed 93% on LoCoMo benchmark

Environment

• OS: Windows 10
• Python: 3.12.7
• Docker: 28.1.1 (MongoDB, Elasticsearch, Milvus, Redis)
• Dependencies: uv sync --group evaluation

Configuration

Models:

• LLM: openai/gpt-4.1-mini (OpenRouter, temp=0.3)
• Embedding: Qwen/Qwen3-Embedding-4B (DeepInfra, dim=1024)
• Reranker: Qwen/Qwen3-Reranker-4B (DeepInfra)
• Search mode: agentic

Commands Run

# Start services
docker-compose up -d

# Smoke test (30 questions, 10 messages/conv)
uv run python -m evaluation.cli --dataset locomo --system evermemos --smoke

# Full conv-26 (152 questions, 419 messages)
uv run python -m evaluation.cli --dataset locomo --system evermemos --from-conv 0 --to-conv 1

Results

Test	Messages	Questions	Accuracy	vs Paper
Paper (LoCoMo)	All	1,986	93.0%	-
Smoke test	10/conv	30	52.22%	-40.78%
Conv-26 (full)	419	152	38.38%	-54.62%

Category Breakdown (Smoke Test)

• Single-hop: 41.67% (vs 96.08% in paper)
• Multi-hop: 54.76% (vs 91.13% in paper)
• Temporal: 66.67% (vs 89.72% in paper)
• Open domain: 100% (vs 70.83% in paper)

Key Findings

1. Performance degrades with more context:

   • 10 messages: 52.22%
   • 419 messages: 38.38% (-13.84%)

2. Only tested 1/10 conversations (7.66% of full benchmark)

3. Possible causes:

   • Retrieval struggles with large memory banks
   • Memory consolidation losing information
   • Different evaluation methodology or configuration

Questions for Authors

Can you share some more details on reproducing the results? Thanks!

[cyfyifanchen]

Hey, thanks for the issue. We just ran the test and everything works as expected.

Do you want to join the Discord channel or WeChat group? We can talk about the way you tested it and make sure you used the right process so you get the same results we did.

[conahpchen]

I have the same config and also get bad test result

[cyfyifanchen]

@conahpchen @wangyu-ustc Are you guys on our Discord or in our WeChat group? If you are, let's collaborate on that. We can arrange a session to talk about it, as I think there's something wrong with the process.

[Malena-yy]

cannot reproduce LoCoMo benchmark results either, any update for this?

[cyfyifanchen]

Closing as part of the EverOS 1.0 issue triage. This issue targets the old benchmark/evaluation code path. The current repo uses the 1.0 server API for LoCoMo reproduction; see docs/locomo_benchmark.md. PR #258 adds migration notes for legacy benchmark reports: #258. If this still reproduces with current main and the 1.0 benchmark commands, please open a fresh issue with the exact command and output.