Cohort memory can support LLM-based recommendation planning through group-level user profiles and advice. As outcome feedback becomes available, this guidance may need to be updated while preserving the underlying observations. We present RecGroupMemory, a workflow that provides this memory to an orchestration LLM and separates memory construction from feedback-based revision. It uses feedback from the previous day to revise decision guidance, while validation checks keep selected source values unchanged. Online comparisons were conducted among Taobao users with low purchasing power selected by the business team. Relative to a no-memory baseline, memory without feedback and memory with feedback showed increases of 1.58% and 1.78%, respectively, in the number of guided transaction users.
Figures & tables
Figure 1. RecGroupMemory system overview. Multi-source evidence and online outcomes support nearline synthesis and feedback refinement. Published Profile and Advise guide the orchestration LLM; stored memory supports subsequent updates. The Evidence Extractor organizes multi-source observations into event, experiment, and context memory in Shallow Memory. The Memory Builder constructs cohort memory and the Memory Refiner revises its guidance; knowledge supports both. Evidence Logs link to Memory Logs by Memory ID and provide outcomes to Memory Refiner. Refined Cohort Memory is published to Shallow Memory, which retains Profile, Advise, and Evidence. Only Profile and Advise feed the orchestration LLM for workflow planning. Online logs return to Evidence Logs, which also update user signals. Stored memory feeds Memory Refiner for feedback refinement in the next cycle.
Figure 2. Feedback-aware revision with at most one repair and original-response fallback. Fixed values are checked against CT . Direction-to-strength checks depend on configuration; semantic policies remain prompt-based. Three outlined stages on a white background show input alignment, decision revision, and validation with writing. Current memory C_T and previous-day feedback F_(T-1) are left-joined by Memory ID and checked for eligibility and source validity. Passing records enter Memory Refiner, supported by optional historical memory H_T from Shallow Memory, knowledge, and prompt policies. Candidate U_T enters validation; an invalid candidate permits one repair and recheck. A valid candidate is accepted. Ineligible or invalid-source records, still-invalid candidates, and handled model failures retain C_T. Accepted and retained responses are written to database partition T with source evidence preserved. Direction-to-strength checks depend on configuration; semantic policies remain prompt-based.
Metric
M/B (E1)
F/B (E2)
Transaction UV
+1.58%
+1.78%
Orders
+0.80%
+2.05%
PCVR
+1.60%
+1.95%
Exposure UV
+0.14%
+0.02%
Click UV
−0.42%
+0.04%
IPV UV
−0.50%
+0.17%
Table 1. Relative changes against each experiment’s no-memory baseline.