Research map / Large language models / Retrieval and knowledge
CAM: A Constructivist View of Agentic Memory for LLM-Based Reading Comprehension
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CAM, a constructivist agentic memory for LLM reading comprehension that builds a hierarchical memory via incremental overlapping clustering (ego-centric node replication plus incremental label propagation) and retrieves with a Prune-and-Grow associative strategy.
Problem
Long-document reading comprehension by LLMs is hampered by limited context and dispersed information; existing agentic memory systems lack a principled design and do not jointly offer structure, flexible many-to-many assimilation, and efficient online/batch updates.
Approach
Grounds memory design in Piaget's constructivist theory (structured schemata, flexible assimilation, dynamic accommodation) and implements it as a hierarchy built by incremental overlapping clustering: ego-network node replication turns overlapping clusters into non-overlapping ones so label propagation can run locally and in batches, followed by LLM summarization of modified clusters and an LLM-g
Key results
- CAM with GPT-4o-mini achieves ACC-L 52.3 on NovelQA and 57.6 on QMSum, outperforming the best baselines (RAPTOR, GraphRAG) by about 3% on average across metrics.
- In batch-level online setting, CAM integrates new chunks about 4x faster than offline RAPTOR and GraphRAG rebuilds at large batch sizes, with stable ACC-L across batch sizes.
- Switching the backbone to open models Llama-3.1-8B-Instruct and Qwen2.5-7B-Instruct yields only slight drops (e.g., NovelQA ACC-L 49.5 and 50.6 vs 52.3).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CAM's constructivist memory design instead of the strongest listed baseline | long-text reading comprehension · NovelQA, QMSum, FABLES, MultiHop-RAG, ODSum-Story, ODSum-Meeting | reported benchmark metrics: +2.3%; +1.7%; +4.5%; +1.4%; +1.3%; +3.7%; +4.3%; +4.0%; +3.4%; +3.9%; +1.8%; +1.2%; +4.4%; +2.2%; +2.7%; +4.9% | Helped |
| add hierarchical clustering to the foundational semantic network (ablation) | CAM · NovelQA, QMSum, MultiHop-RAG, ODSum-Story | ACC-L / EM: 52.3 -> 46.7; 57.6 -> 46.6; 72.8 -> 68.5; 54.6 -> 49.3 | Helped |
| add ego-centric disentanglement for flexible overlapping clustering (ablation) | CAM · NovelQA, QMSum, MultiHop-RAG, ODSum-Story | ACC-L / EM: 52.3 -> 50.3; 57.6 -> 51.5; 72.8 -> 70.3; 54.6 -> 51.2 | Helped |
| add fine-grained foundational modeling (ablation) | CAM · HotpotQA, 2WikiMultiHopQA, MuSiQue | EM and F1: 60.7 -> 62.1; 75.4 -> 76.5; 61.4 -> 65.8; 70.2 -> 75.3; 24.6 -> 26.8; 37.2 -> 38.5 | Helped |
| add fine-grained foundational modeling (ablation) | CAM · NovelQA, QMSum, MultiHop-RAG, ODSum-Story | ACC-L / EM: 52.3 -> 53.5; 57.6 -> 57.2; 72.8 -> 72.0; 54.6 -> 54.8 | Mixed |
| replace Prune-and-Grow with hierarchical traversal or global retrieval (ablation) | CAM · NovelQA, QMSum, MultiHop-RAG, ODSum-Story | ACC-L / EM: 52.3 -> 46.5; 52.3 -> 50.3; 57.6 -> 55.2; 57.6 -> 56.3; 72.8 -> 68.3; 72.8 -> 70.5; 54.6 -> 49.5; 54.6 -> 50.8 | Hurt |
| replace GPT-4o-mini with Llama-3.1-8B-Instruct or Qwen2.5-7B-Instruct (ablation) | CAM · NovelQA, QMSum, MultiHop-RAG, ODSum-Story | R-L / ACC-L / EM / F1: 25.4 -> 23.7; 52.3 -> 49.5; 26.5 -> 25.3; 57.6 -> 54.4; 72.8 -> 70.3; 77.5 -> 74.2; 25.5 -> 24.3; 54.6 -> 52.5; 25.4 -> | Hurt |
| replace text-embedding-3-small with text-embedding-3-large or E5-Mistral-7B-Instruct (ablation) | CAM · NovelQA, QMSum, MultiHop-RAG, ODSum-Story | R-L / ACC-L / EM / F1: 25.4 -> 25.8; 52.3 -> 53.0; 26.5 -> 27.4; 57.6 -> 59.2; 72.8 -> 73.2; 77.5 -> 78.0; 25.5 -> 26.7; 54.6 -> 56.4; 25.4 -> | Helped |
Surprising findings
- Fine-grained LLM knowledge-graph extraction (as in GraphRAG/HippoRAG) increases cost more than threefold without commensurate gains on narrative tasks.
- MemTree's chunk-by-chunk online insertion becomes slower than offline reconstruction when the new batch exceeds about 400 chunks.
Limitations
- Relies on fixed prompts and manually tuned hyperparameters rather than learned memory policies.
- Assumes source texts are internally consistent; does not handle contradictory information.
- LLM-based summarization may propagate hallucinations up the hierarchy.
Experimental setup
Datasets: NovelQA QMSum FABLES MultiHop-RAG ODSum-Story ODSum-Meeting HotpotQA MuSiQue 2WikiMultiHopQA
Benchmarks: NovelQA QMSum FABLES MultiHop-RAG ODSum-Story ODSum-Meeting HotpotQA MuSiQue 2WikiMultiHopQA
Metrics: ROUGE-1 ROUGE-L ACC-L exact match F1 F1 positive F1 negative
Baselines: FullContext MemGPT ReadAgent RAPTOR GraphRAG HippoRAG MemTree
Builds on: RAPTOR MemTree GraphRAG HippoRAG
Closest papers by meaning
- A Human-Inspired Reading Agent with Gist Memory of Very Long Contexts (ICML 2024 · similarity 0.89)
- Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (ICLR 2026 · similarity 0.89)
- A-Mem: Agentic Memory for LLM Agents (NeurIPS 2025 · similarity 0.88)
- Harness the Memory: A Holistic Evaluation of Memory Substrates in Memory Agents (NeurIPS 2026 · similarity 0.87)
- REMem: Reasoning with Episodic Memory in Language Agent (ICLR 2026 · similarity 0.87)
- MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.