Layered Convergence in Autonomous Agent Memory: A Multi-Model Cognitive Architecture for Persistent Recall in Long-Running LLM Agents
Open accessThe DOI matches its Crossref record.
Abstract
Autonomous agents built on large language models (LLMs) operate under a structural limitation that constrains their utility in sustained, multi-domain environments: each session begins with no episodic, procedural, or declarative memory of prior sessions. Existing memory augmentation systems address this limitation through single-model approaches, including vector-based extraction (Mem0), temporal knowledge graphs (Zep/Graphiti), multi-strategy retrieval (Hindsight), OS-inspired tiered management (Letta/MemGPT), and framework-coupled recall (LangMem), each optimizing for one category of recall while leaving others unaddressed. This paper presents a multi-model cognitive architecture in which five complementary memory systems operate concurrently: (1) identity-prompt memory providing behavioral continuity via session-loaded configuration files; (2) structured constraint memory enabling programmatic policy enforcement with audit trails; (3) vector-enhanced semantic memory combining BM25 keyword matching with cosine-similarity search, temporal decay, and maximal marginal relevance re-ranking; (4) relational learning memory closing feedback loops between expert recommendations and their real-world outcomes; and (5) a hierarchical temporal knowledge graph organized into domain-scoped partitions with cross-domain tunnels and dynamic context assembly. Each model was introduced to address a specific failure mode observed during eight weeks of production deployment, and each operates additively without replacing its predecessors. A combinatorial ablation study across all 32 possible subsets of the five models, totaling 3,520 evaluation runs across eight capability dimensions, validates the architecture using Shapley value decomposition. Three findings stand out. First, only two of the five models carry measurable weight under structural evaluation: Model II accounts for 104% of constraint enforcement capability, and Model V accounts for 100% of all retrieval capability, while Models I, III, and IV show near-zero contribution, an expected result given the evaluation's routing through constraint-check and palace-search pathways. Second, upgrading the constraint checker from keyword matching to semantic rule detection raised constraint enforcement from 59.8% to 73.3%, with Model II's Shapley value increasing from 0.286 to 0.430. Third, two new evaluation dimensions reveal previously unmeasured contributions: semantic recall scenarios show Model III (vector search) providing independent recall value (Shapley 0.025, 25% of total), and cross-model flow scenarios confirm that Models II and V contribute additively with zero interaction. The full architecture achieves 80.0% on factual recall with zero policy violations observed across nine weekly advisory council cycles in production. Constraint enforcement, feedback-loop learning, and dynamic context assembly are not present in any evaluated commercial or open-source memory system. The architecture was deployed without modifying the host platform's source code, suggesting portability across agent frameworks.
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
Citations
Metadata
- Authors
- Harris, L.
- Corresponding author
- Harris, Lance
- Venue
- SSRN (preprint)
- Publisher
- Elsevier BV
- Published
- 2026
- ISSN
- 1556-5068 (ISSN-L)
- OpenAlex
- W7155561645
- Cited by
- 0
- References
- 0
- Type
- posted content
- Language
- en
- Open access
- green
- doi.org
- Locations
- SSRN Electronic Journal (acceptedVersion)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › AI-based Problem Solving and Planning
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Information Systems › Information Retrieval and Search Behavior
- Keywords
- LLM agents, autonomous agents, persistent memory, session continuity, constraint enforcement, vector search, semantic memory, temporal knowledge graphs, hierarchical knowledge graph, Shapley decomposition, ablation study, long-term recall, policy violation detection, cross-domain knowledge transfer, retrieval-augmented generation
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.