CoChat

Context Collapse in Long-Horizon Agents: Benchmarking Hierarchical Memory against RAG and Summarization

Open access
Raheem, Ebaad · 2026 · SSRN Electronic Journal

The DOI matches its Crossref record; 9 of the 9 references it cites were found in a registry or an index.

Abstract

Autonomous Large Language Model (LLM) agents that operate over long conversations face a critical weakness: context collapse, where factual details gradually drift and similar meanings interfere, weakening the agent's ability to track its own state. This letter presents a carefully controlled benchmark of four memory designs; Baseline (full, untruncated history), Rolling Summary (repeated summarization of the dialogue), Retrieval-Augmented Generation (RAG), and a Hierarchical Memory system; evaluated across 800 fully reproducible conversational runs covering Legal, Medical, Tech, and Travel domains. Key facts are inserted at turn one and queried at five fixed recall distances using a two-level evaluation method that prevents the judge model from hallucinating. The Hierarchical architecture reaches an overall Factual Retention Rate (FRR) of 41.6%, performing competitively with the Baseline (40.6%) while using 34.8% fewer context tokens. RAG and Rolling Summary achieve 37.9% and 31.0%, respectively. Rolling Summary reduces context tokens by 60.5% but raises LLM API calls by roughly 205%, making it both expensive and harmful to performance. A strong primacy effect appears across all architectures as the conversation grows: at medium lengths (distances 14-19) retention becomes erratic (Baseline FRR falls to 39.0%), but then retention paradoxically jumps to 55.0% at maximum length, as attention locks tightly onto the start of the prompt. Domains differ sharply in fragility; the Medical domain collapses to a Baseline FRR of 12.0% and generates 192 explicit refusals under Rolling Summary, whereas Legal sustains 88.8%. These findings show that even with the full conversation history available, the Baseline essentially ties with hierarchical memory (40.6% vs. 41.6%, p = 0.31), yet domain collapse (Medical 12.0%) remains. Full context alone is not enough; memory structured at multiple levels offers greater resilience.

Overview

Integrity signals

Registry

Crossref · 10.2139/ssrn.6976218

Cites

9 of 9 references verified0 citation errors · 0 not checkable

Statistics

Recomputing p-values needs the paper’s full text.

Verification

Claims the paper makes

We couldn’t retrieve this paper’s full text, so its claims and self-check weren’t run. The registry and citation checks still apply.

Citations

Cites

Show all 9 references

Research works printed but not linked: 1

  1. Sung Joon, 2023, Generative agents: Interactive simulacra of human behavior, Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST), 1

Metadata

Authors
Raheem, E.
Corresponding author
Raheem, Ebaad
Venue
SSRN Electronic Journal
Publisher
Elsevier BV
Published
2026
ISSN
1556-5068 (ISSN-L)
OpenAlex
W7167492492
Cited by
0
References
4
Type
posted content
Language
en
Open access
Locations
SSRN Electronic Journal (acceptedVersion)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Speech and dialogue systems
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › AI-based Problem Solving and Planning
Keywords
context collapse, hierarchical memory, retrieval-augmented generation, primacy effect, chatbots

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.