CoChat

The DOI matches its Crossref record.

Beyond Retrieval vs Context: A Unified Evaluation Framework for External Information Management in LLM Agents

Open access
Rocchi, Alessio · 2026 · SSRN

Abstract

Recent advances in language-model context windows past one million tokens have reopened the debate over whether retrieval-augmented generation remains necessary. We argue that the debate is misframed: long-context inference, retrieval-augmented generation (RAG), and agent memory systems are not paradigms in competition but three architectural realisations of a single information-management policy. We introduce a five-axis framework-accuracy, cost, latency, freshness, traceability-that scores any policy under a workload distribution and a price schedule. A four-operator canonical form (ingest, consolidate, forget, working-tier maintenance) reduces memory-augmented generation to RAG composed with three or four additions, and a precise retrieval-equivalence relation identifies the conditions under which a deployed memory system collapses to RAG observationally. We exhibit this collapse empirically on the open-source Mem0 2.x release. An engineering map of six surveyed conversationalmemory systems (Mem0, Letta, Zep, A-MEM, Cognee, Generative Agents) documents four cross-system patterns: forgetting is uniformly absent or opt-in, consolidation is heterogeneous in scheduling, papervs-code divergence runs in both directions, and the cognitive memory taxonomy lives in documentation rather than in schemas. A controlled empirical protocol (§6) is preregistered at horizons of 10 2 , 10 3 , and 10 4 turns under an explicit and frozen cost function; the present draft executes the 10 2-turn…

Overview

Integrity signals

Registry

Crossref · 10.2139/ssrn.6830898

Cites

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

Metadata

Authors
Rocchi, A.
Corresponding author
Rocchi, Alessio
Venue
SSRN
Publisher
Elsevier BV
Published
2026
ISSN
1556-5068 (ISSN-L)
OpenAlex
W7164017697
Cited by
0
References
0
Type
posted content
Open access
Locations
SSRN Electronic Journal (acceptedVersion)
Topics
  • Social Sciences › Cultural Studies › Language and cultural evolution
  • Physical Sciences › Computer Science › Artificial Intelligence › Multi-Agent Systems and Negotiation
  • Physical Sciences › Computer Science › Artificial Intelligence › Speech and dialogue systems
Keywords
retrieval-augmented generation, long-context inference, LLM agents, memory consolidation, forgetting, evaluation framework

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.