MemoryCPT: An End-to-End Agent Memory Framework for Cost-Performance Trade-off¶
Core memory operator mini-suite, not paper-scale agent evaluation.
Paper information¶
| Field | Details |
|---|---|
| Paper | Paper or original source |
| Institution | Hong Kong University of Science and Technology / Tencent LIGHTSPEED STUDIOS. |
| First public release | 2026-08-05 (arXiv v1) |
| Author code | No public author code found as of 2026-08-09. |
| Local adapter or method key | memorycpt |
| Local reproduction code | src/auto_research/agent_research/latest_20260809.py |
Paper method and evidence¶
Distill offline memory construction into a compact QAD model. QAR retrieves via RRF and summarizes query-relevant context with LoRA; cost-aware GRPO optimizes quality per cost.
flowchart LR
A[Distill memory construction] --> B[Retrieve with RRF]
B[Retrieve with RRF] --> C[Summarize query-relevant memory]
C[Summarize query-relevant memory] --> D[Optimize quality per cost]Original paper figure¶
Figure shows memory construction, retrieval and summarization. From the original paper; copyright remains with the authors.
Core equation¶
\(QPC=Q(answer)/C_{inference}\).
Reported results¶
Paper reports improved quality–cost tradeoffs on LoCoMo and LongMemEval, with ablations for QAD, RRF, QAR and cost reward.
Local reproduction¶
Local metrics. Same mini-suite and mechanism-disabled comparator; no author model, proprietary environment or expensive judge. Evolve can discover the operator, but genome composition is constrained by public evaluation and budget. Checkpoints and raw runs are not committed.