Building agent taste from logged attention: memory and pipeline stacks
apocalyptic_cc · reddit · 2026-08-15
The author discusses building Personal Agent taste based on actual user interaction signals (clicks, dwell time, saves, returns) rather than stuffing context with everything. Four technical approaches are weighed: a weighted memory store boosting content by dwell time, fine-tuning on dwell-time-weighted excerpts, using engagement as a RAG retrieval signal, and explicit preference models scored from behavior. The author seeks to compare technical stacks and concrete examples of agents surfacing better results via attention data.
More from coding & agent
- Atomic + Qwen 3.8 27B benchmarks show 2x speedup over Hermes and Prime — SucceededMind · 2026-08-15
- Consilience: Local-first ontology workspace with unified agent — aurorax0 · 2026-08-15
- Mermail Launches Email and Wallet for AI Agents to Solve Online Payments — Shruti_0810 · 2026-08-15
- Analysis on DeepSeek's choice of Loopx and their development velocity — teortaxesTex · 2026-08-15
- Open-Source Project Replicates Full LLM Training Pipeline in Pure PyTorch — thisguyknowsai · 2026-08-15
- GLM-5.3 Review: Matches GPT-4 Coding, and I Built 3 Plugins with It — 赛博禅心 · 2026-08-15