AuthorsXin Zhang, Kailai Yang, Chenyue Li et al.
TL;DR
MemAdapter uses generative subgraph retrieval from a unified memory space to align heterogeneous agent memories, reaching F1 61.59 on NarrativeQA with Qwen2.5-7B versus 56.91 for MemoryLLM (+4.68).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Existing agent memory systems tightly couple retrieval to specific paradigms like explicit, parametric, or latent memory, preventing cross-paradigm fusion and reuse.
Graph-based systems adapted to latent memory require extensive architectural redesign and continual training, suffering from catastrophic forgetting that degrades original retrieval and blocks effective memory fusion.
HOW IT WORKS
MemAdapter centers on a Generative Subgraph Retriever, Anchored Alignment Module, Target Alignment Module, Unified Memory Space, and Agent Model to unify heterogeneous memory paradigms.
You can think of MemAdapter as a universal memory bus: different memory types plug in via small adapters, and the bus feeds a graph-shaped RAM of task-specific evidence.
This design lets MemAdapter generate structured evidence subgraphs from any aligned memory state, enabling flexible, interpretable retrieval far beyond what a plain context window can support.
DIAGRAM
This diagram shows how MemAdapter first distills a generative subgraph retriever on an anchored paradigm and then trains lightweight alignment modules via contrastive learning.
DIAGRAM
This diagram shows how MemAdapter is evaluated on three QA benchmarks with multiple agent scales and baseline memory systems.
PROCESS
Memory State Construction
MemAdapter uses MemBuild to turn long-horizon memory D into a structured Memory State M under explicit, parametric, or latent paradigms.
Generative Subgraph Retrieval via Model Distillation
MemAdapter trains the Generative Subgraph Retriever on an anchored graph-based paradigm, imitating a teacher that provides query-conditioned evidence subgraphs.
Cross-Paradigm Alignment with Contrastive Learning
MemAdapter optimizes a Target Alignment Module so that target memory states map into the Unified Memory Space consistent with the anchored alignment.
Plug-and-Play Memory Retrieval and Fusion
MemAdapter fuses aligned representations from multiple paradigms, generates an explicit evidence subgraph, and passes it to the Agent Model to answer the query.
KEY CONTRIBUTIONS
Unified generative subgraph retrieval framework
MemAdapter introduces a Generative Subgraph Retriever over a Unified Memory Space, trained via distillation on 40,000 HotpotQA instances using Qwen2.5-1.5B with LoRA.
Fast cross-paradigm alignment modules
MemAdapter trains lightweight Anchored Alignment Module and Target Alignment Module via contrastive learning on only 2,500 demonstrations, finishing alignment within 13 minutes on a single GPU.
Zero-shot heterogeneous memory fusion
MemAdapter projects multiple memory states into the Unified Memory Space and fuses them via max-pooling, enabling plug-and-play fusion of explicit, parametric, and latent memory for the same agent.
RESULTS
F1
61.59
+4.68 over MemoryLLM on NarrativeQA with Qwen2.5-7B
ROUGE-1
63.22
+4.49 over MemoryLLM on NarrativeQA with Qwen2.5-7B
Average F1
43.84
+7.56 over StreamingLLM on three benchmarks with Qwen2.5-7B
Memory Length
2241.80
MemAdapter uses less than half the characters of StreamingLLM's 5884.60
These numbers come from WikiMultiHopQA, NarrativeQA, and MuSiQue, which stress multi-hop and long-horizon reasoning. MemAdapter's gains show that generative subgraph retrieval plus fast alignment can beat paradigm-specific systems while compressing memory and improving utilization.
BENCHMARK
These numbers come from WikiMultiHopQA, NarrativeQA, and MuSiQue, which stress multi-hop and long-horizon reasoning. MemAdapter's gains show that generative subgraph retrieval plus fast alignment can beat paradigm-specific systems while compressing memory and improving utilization.
BENCHMARK
F1 on NarrativeQA comparing MemAdapter to baseline memory systems.
KEY INSIGHT
MemAdapter cuts average memory length to 2241.80 characters yet reaches memory utilization up to 61.88%, higher than all baselines at 7B scale.
This is surprising because parametric and latent systems process over 5884.60 characters, but storing more information does not translate into better utilization or downstream accuracy.
WHY IT MATTERS
MemAdapter unlocks a unified way to retrieve explicit evidence subgraphs from any agent memory paradigm using small alignment modules instead of paradigm-specific retrievers.
Builders can now plug diverse memory systems into one agent, align them in under 13 minutes, and perform zero-shot fusion of explicit, parametric, and latent memory without retraining the agent.
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…