Agentic memory retrieval: how to do variable top-k in low-latency tier
WonderfulArt9908 · reddit · 2026-09-05
- A team building enterprise agentic memory (highly curated, few-thousand-entry knowledge snippets per agent) shares their three-tier retrieval setup: Latency, Balanced, Accuracy.
- Their low-latency tier uses BM25 + dense embedding → top-50 → rerank → top-25, but they doubt it.
- Core open question: how to do principled variable top-k — RRF scores can't be thresholded, and embedding distance distributions differ wildly across agent domains (Text2SQL vs legal), making fixed thresholds unreliable.
More from coding & agent
- Dev patches WebKitGTK CVE to restore drag-and-drop, ships Copilot app as Flatpak — unixterminal · 2026-09-05
- Nous Research's Hermes Desktop lets you build fleets of specialist agents that talk to each other — Teknium · 2026-09-05
- Skip the Chain-of-Thought: Monitor Agent Inputs and Outputs Instead — nicolascraske · 2026-09-05
- Voice agent testing tools compared: Cekura, Cyara and TestMu solve different problems — Fishful_Revenge · 2026-09-05
- LinkedIn VP Breaks Down Hiring Assistant: The Problem Is Fragmented Recruiting Workflows — CodeByPoonam · 2026-09-05
- Agent frameworks are 10% of the problem—state hygiene is what breaks production systems — Deepfeet-09 · 2026-09-05