AuthorsSongxin Lei, Kun Ouyang, Weilin Ruan et al.
TL;DR
MemoryCPT uses Query-agnostic Distillation plus cost-aware Query-aware Retrieval and Summarization to reach F1 0.479 on LoCoMo with Cost 4.31 (QPC 0.111).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Long-horizon LLM agents must recover useful evidence from large histories, but existing memory pipelines introduce redundant context and high inference cost.
These training-free systems rely on hand-crafted heuristics and repeated LLM calls, making multi-session dialogue QA expensive while only a small fraction of tokens contain useful evidence.
HOW IT WORKS
MemoryCPT’s core mechanism is Query-agnostic Distillation, Query-aware Retrieval and Summarization, RRF-based Retrieval, and a GRPO-trained Fine Summary adapter over a frozen QA model.
You can think of MemoryCPT like RAM and disk: QAD builds structured long-term storage, while QAR acts as a smart memory controller deciding which blocks to load.
This KEY_MECHANISM lets MemoryCPT learn when to construct, retrieve, and compress memories, achieving cost-aware context selection that a plain context window cannot match.
DIAGRAM
This diagram shows how MemoryCPT processes a query online using RRF-based retrieval and GRPO-trained summarization before calling the frozen QA model.
DIAGRAM
This diagram shows how MemoryCPT trains Query-agnostic Distillation and Query-aware Retrieval and Summarization to optimize Quality per Cost.
PROCESS
Query-agnostic Distillation
MemoryCPT runs Query-agnostic Distillation offline, where teacher traces train LoRA-A to perform BatchSegmenter, EpisodeGenerator, EpisodeMerger, and SemanticExtractor operations.
Memory Store Construction
MemoryCPT merges LoRA-A into the base model to form baseA, then builds episodic and semantic memories M = Mep ∪ Msem from raw histories H.
Query-aware Retrieval and Summarization
For each query q, MemoryCPT applies RRF-based Retrieval over M to get MRRF q and uses the GRPO-trained Fine Summary adapter LoRA-B to generate a compact mq.
Final QA with Cost-aware Context
MemoryCPT feeds q and mq into the frozen Qwen3-14B QA model, producing answer y^ while controlling token cost to maximize Quality per Cost.
KEY CONTRIBUTIONS
End-to-end memory pipeline for cost-performance optimization
MemoryCPT spans Query-agnostic Distillation and Query-aware Retrieval and Summarization, explicitly optimizing Quality per Cost on long-horizon multi-turn dialogue memory.
Two-stage post-training algorithm
MemoryCPT trains LoRA-A via supervised fine-tuning on teacher reasoning traces, then trains LoRA-B with GRPO and a cost-aware reward while keeping the QA model frozen.
Empirical evaluation of Quality per Cost
MemoryCPT introduces QPC and shows F1 0.479 with Cost 4.31 on LoCoMo, improving QPC over BudgetMem from 0.015 to 0.111 with Qwen-2.5-7B.
RESULTS
F1
0.479
+0.106 over BudgetMem on LoCoMo with Qwen-2.5-7B
Judge
0.755
+0.115 over BudgetMem on LoCoMo with Qwen-2.5-7B
Cost
4.31 (USD×10^4 per query)
-19.80 vs BudgetMem’s 24.11 on LoCoMo with Qwen-2.5-7B
QPC
0.111
+0.096 vs No-Memory Qwen-3-14B’s 0.015 on LoCoMo
On LoCoMo, which tests long-horizon dialogue memory, MemoryCPT with Qwen-2.5-7B base achieves F1 0.479 and Cost 4.31, yielding QPC 0.111. This MAIN_RESULT shows MemoryCPT can increase answer quality while cutting inference cost compared to BudgetMem and other baselines.
BENCHMARK
On LoCoMo, which tests long-horizon dialogue memory, MemoryCPT with Qwen-2.5-7B base achieves F1 0.479 and Cost 4.31, yielding QPC 0.111. This MAIN_RESULT shows MemoryCPT can increase answer quality while cutting inference cost compared to BudgetMem and other baselines.
BENCHMARK
F1 on LoCoMo for MemoryCPT and representative memory baselines using Qwen-2.5-7B as the base.
BENCHMARK
Inference Cost (USD×10^4 per query) on LoCoMo for MemoryCPT and baselines.
KEY INSIGHT
MemoryCPT with Fine Summary reaches F1 0.479 on LoCoMo at Cost 4.31, while a variant without FS has F1 0.446 at Cost 8.35.
It is surprising that adding a GRPO-trained summarizer both improves quality by 0.033 and cuts cost by 4.04, contradicting the intuition that better summaries must be more expensive.
WHY IT MATTERS
MemoryCPT unlocks long-horizon agents that can learn memory construction and query-aware compression jointly, targeting Quality per Cost instead of raw accuracy.
Builders can now deploy agents that maintain rich episodic and semantic memories while keeping inference budgets low, making multi-session personalization practical at scale.
Peilin Feng, Suorong Yang, Soujanya Poria
arXiv 2026 · 2026
Σ-Mem maintains historical competence evidence matrices M_p, a peer relationship matrix G, a memory readout r_{p,t}, and residual steering δ_{p,t} to guide multi-agent trust decisions. On the mixed counterfactual benchmark, Σ-Mem raises Qwen3-0.6B accuracy at CF@90 from 46.22% with the base model to 71.10% with Σ w/ G.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Answers use this explainer on Memory Papers.
Checking…