Question on Phase III latency + ablation without Agentic Verification/Query Rewriting

[BaoBaoGitHub]

Hi EverMemOS team,

Thanks for sharing EverMemOS!

I have a question specifically about Phase III (Reconstructive Recollection). As described, it includes:

1. MemScene Selection
2. Episode and Foresight Filtering
3. Agentic Verification and Query Rewriting

The third step (Agentic Verification and Query Rewriting) may trigger multiple LLM calls (e.g., sufficiency check + query rewriting + additional retrieval rounds). In some real-time conversational systems, memory retrieval latency is tightly constrained (e.g., staying within ~200ms on the retrieval path), so I’m trying to understand the runtime implications of this module.

Could you please share:

1) Retrieval latency measurements (if available)

• Did you measure end-to-end retrieval latency for Phase III (p50/p95)?

2) Ablation results without Agentic Verification and Query Rewriting

• Did you run an ablation where Phase III is “one-shot retrieval” (MemScene Selection + Episode/Foresight Filtering only), i.e., removing the verification/rewriting loop entirely?
• If yes, could you share the performance impact on LoCoMo/LongMemEval/PersonaMem v2, and any notes on the trade-off between accuracy and latency?

This would be very helpful for practitioners who want to adopt EverMemOS in latency-sensitive settings.

Thanks in advance!

[freshpomelo]

Hi BaoBao, we have several query method. If you have a tightly constrained retrieval latency requirement, you can use keyword (BM25), vector similarity, and hybrid search without agentic search. We did a test in our cloud env and the performance for search API without agentic search is P99: 606ms P50: 137.33ms.
About your 2nd question, we need to do further experiment.

[BaoBaoGitHub]

Hi BaoBao, we have several query method. If you have a tightly constrained retrieval latency requirement, you can use keyword (BM25), vector similarity, and hybrid search without agentic search. We did a test in our cloud env and the performance for search API without agentic search is P99: 606ms P50: 137.33ms. About your 2nd question, we need to do further experiment.

Hi.

Could you share whether there are any recent results on retrieval latency and performance when removing query rewriting in agentic mode?