AuthorsRui Li, Zeyu Zhang, Xiaohe Bo et al.
TL;DR
CAM uses an incremental overlapping clustering algorithm with Prune and Grow retrieval to reach 52.3 ACC-L on NovelQA, +4.5 over RAPTOR.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
CAM targets scenarios where LLM performance declines as input texts lengthen, even within the specified context length.
When long novels, meetings, or multi document corpora exceed context, reading agents miss dispersed evidence, harming question answering, summarization, and claim verification.
HOW IT WORKS
CAM centers on an incremental overlapping clustering algorithm, Foundational Semantic Network G0, and Ego Centric Disentanglement to build hierarchical schemata.
You can think of CAM like a human schemata system plus a dynamic card catalog, where chunks are filed into multiple overlapping folders that are locally rebalanced.
This design lets CAM flexibly assimilate new chunks and dynamically accommodate structure changes, enabling retrieval patterns that a flat context window or static tree cannot support.
DIAGRAM
This diagram shows how CAM performs Prune and Grow associative exploration over its memory hierarchy when answering a query.
DIAGRAM
This diagram shows how CAM is evaluated across offline and batch level online settings on long text benchmarks.
PROCESS
Foundational Network Expansion
CAM encodes new contiguous text chunks into the Foundational Semantic Network G0 using a composite similarity score with femb and positional proximity.
Ego Centric Disentanglement
CAM extracts ego networks, partitions them into connected components, and builds a Replica Network that disentangles overlapping clusters via node replication.
Online Clustering Updates
CAM runs an incremental label propagation algorithm on the Replica Network, updating clusters and regenerating abstraction nodes in higher level schemata.
Prune and Grow Associative Retrieval
Given a query, CAM performs fast localization with global similarity, then iteratively expands and prunes activated nodes to feed the LLM backbone for answers.
KEY CONTRIBUTIONS
Constructivist design principle for agentic memory
CAM formalizes structured schemata, flexible assimilation, and dynamic accommodation using an incremental overlapping clustering algorithm and ego centric disentanglement for long text reading.
Prototype of Constructivist Agentic Memory
CAM instantiates the principle with a Foundational Semantic Network, Replica Network, and Prune and Grow retrieval, supporting both offline and batch level online memory development.
Evaluation on diverse long text benchmarks
CAM achieves dual superiority in performance and efficiency, e.g., 52.3 ACC L on NovelQA and over 4× faster batch insertion than RAPTOR and GraphRAG.
RESULTS
ACC-L
52.3
+4.5 over RAPTOR on NovelQA
R-L
25.4
+1.7 over RAPTOR on NovelQA
ACC-L
57.6
+3.7 over GraphRAG on QMSum
ACC-L
54.6
+4.4 over GraphRAG on ODSum Story
Table 2 reports CAM on NovelQA, QMSum, and ODSum Story, which test long form question answering and summarization. The gains in ACC L and ROUGE show that CAM’s constructivist memory design improves both answer quality and abstraction over strong structured baselines like RAPTOR and GraphRAG.
BENCHMARK
Table 2 reports CAM on NovelQA, QMSum, and ODSum Story, which test long form question answering and summarization. The gains in ACC L and ROUGE show that CAM’s constructivist memory design improves both answer quality and abstraction over strong structured baselines like RAPTOR and GraphRAG.
BENCHMARK
ACC-L on NovelQA single document question answering comparing CAM against structured and unstructured memory baselines.
KEY INSIGHT
CAM’s online batch insertion becomes over 4× faster than RAPTOR and GraphRAG while maintaining ACC L comparable to the offline setting.
This is surprising because many expect dynamic online memory updates to be slower and less stable than a single offline reconstruction over the full corpus.
WHY IT MATTERS
CAM unlocks constructivist style agentic memory where schemata flexibly assimilate new chunks and dynamically accommodate structure without full rebuilding.
Builders can now deploy long context reading agents that continuously ingest streaming chapters or news batches while preserving retrieval quality and keeping latency practical.
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%.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Answers use this explainer on Memory Papers.
Checking…