Memory-Orchestrated Semantic System (MOSS): An Auditable Agentic Memory Architecture

Abstract: Long-term memory remains a structural weakness of AI agents. The dominant approach, retrieval-augmented generation (RAG), relies on embedding-based similarity search, which is opaque by construction, difficult to audit, and bounded by the theoretical limits of vector representations. We present the Memory-Orchestrated Semantic System (MOSS), an agentic memory architecture in which the agent drives retrieval over a structured relational database. MOSS is model-agnostic, storage-agnostic, and API-agnostic: it runs on any relational engine, connects to any LLM provider (or to deterministic non-LLM processes), and deploys on any infrastructure, local or cloud. Its retrieval execution is symbolic and reproducible (once a query is formulated, no LLM participates in the retrieval loop) and every step of the system, from indexing to answer formulation, is logged and inspectable, making MOSS auditable by construction. Rather than imposing an external ontology, MOSS derives its conceptual vocabulary from the corpus itself. We report on a longitudinal deployment unique in the agentic-memory literature: a year of continuous production over an individual scholar's working corpus--a conversational corpus reaching back to October 2024 (some 44 million tokens, retroactively indexed) comprising 110,183 segments, alongside 163,494 catalogued documents, 569 inductively derived concepts, 322,662 concept annotations, and eleven metadata graphs totaling approximately five million relations--across four successive infrastructure generations. While the present case is that of a single researcher, the architecture is in no way specific to one person: it serves a team, an institution, or any entity that accumulates knowledge over time. We argue that auditable, sovereign, structurally unbounded memory is a precondition for AI agents intended to accompany a person or an organization over years rather than sessions.

• The paper introduces MOSS as an alternative to vector-based retrieval, employing explicit SQL execution to ensure user-auditable memory access.
• It features a triple-agnostic design that decouples model, storage, and execution, enabling seamless portability across various infrastructures.
• A year-long deployment validates MOSS’s effectiveness with scalable inductive ontology, efficient affective indexing, and transparent query trails.

Memory-Orchestrated Semantic System (MOSS): An Auditable Agentic Memory Architecture

Overview and Motivation

The paper presents MOSS, a robust agentic memory architecture designed to overcome the limitations of classical vector-based RAG approaches. The motivation is rooted in the need for stateless LLM agents to have auditable, structurally unbounded memory that supports continuity across long temporal spans. MOSS is distinct in its model, storage, and API agnosticism—it operates on any relational engine, connects to any LLM or deterministic agent, and is deployable across any infrastructure. Crucially, retrieval is symbolic and deterministic, with explicit SQL execution and end-to-end audit trails, departing from opaque, embedding-driven pipelines.

Positioning within Contemporary Agent Memory Architectures

The authors systematically situate MOSS among four methodological families:

• Vector RAG: MOSS challenges RAG’s opacity, irreproducibility, and theoretical bounds (S et al., 31 Mar 2026, Weller et al., 28 Aug 2025), asserting that similarity-based retrieval is fundamentally unfit for personal or regulated long-term memory applications. MOSS avoids single-vector dependence, decouples retrieval from embedding models, and eliminates geometric proxies for relevance in favor of symbolic, explainable queries.
• GraphRAG: While entity-relation graphs add structure and temporality (Peng et al., 2024, Zhang et al., 21 Jan 2025, Yang et al., 5 Feb 2026), MOSS maintains the territory (raw corpus) intact, using graphs as overlays and navigation layers rather than lossy extracts. Temporal adjacency and affective resonance are indexed natively, exceeding common graph-memory practices.
• Production Memory Layers: In contrast to hybrid or proprietary engines (e.g., Mem0 (Chhikara et al., 28 Apr 2025), Zep (Rasmussen et al., 20 Jan 2025), Dreaming [OpenAI, 2026]), MOSS guarantees sovereignty and auditability by decoupling agentic formulation from symbolic execution. No synthesis is opaque; all memory is user-owned, transparent, and portable.
• Agentic Local-SQL Memory: MOSS matches and extends the movement toward structured, locally-owned, SQL-native memory (Biswal et al., 22 Jan 2026, Lin et al., 3 Apr 2026, Bhardwaj, 17 Feb 2026). It distinguishes itself via scale (lifetime corpus), ontology induction from corpus, query-time relevance weighting, affective indexing, and unprecedented longitudinal operational evidence.

Architectural Commitments and System Design

Three fundamental principles anchor MOSS:

• Structural Auditability: All retrieval events are explicit SQL queries, logged in real time. The corpus of record is preserved in human-readable plain text, avoiding proprietary formats and latent spaces.
• Triple Agnosticism: The system is decoupled from model, storage, and execution backend, surviving migrations across multiple infrastructures and LLM providers, as empirically validated in four production generations.
• Inductive Ontology: A concept vocabulary—569 concepts applied in 322,662 segment annotations—is derived bottom-up. This mirrors codebook thematic analysis, enabling domain-specific, corpus-fitting ontologies instead of imported schemas or distributional entity extraction.

MOSS orchestrates agentic query profiling via a dedicated layer which parameterizes SQL queries according to intent (temporal, thematic, affective, etc.), separating stochastic agentic formulation from deterministic database execution. Segments (110,183) are annotated with summaries, affective state, and concepts; documents (163,494) are indexed with summaries and hierarchical structural outlines for efficient partial reading.

The relational map is enriched with eleven metadata graphs totalling five million relations, connecting segments via co-occurrence, temporal adjacency, thematic cohesion, and affective resonance. The territory remains intact, with overlays (calques) tracing semantic, conceptual, and affective strata for agent retrieval.

Longitudinal Deployment Insights

A year-long production deployment validates the architecture’s claims, uniquely among agentic memory systems. The system has continuously supported an active scholar's research, teaching, and organizational activities over a corpus spanning 44 million tokens and 600 conversation files. The architecture survived multiple infrastructure transitions (consumer cloud, Azure Data Lake, local workstation, dedicated VPS) and numerous LLM models (GPT-4o, Gemini, Claude), demonstrating portability and durability.

Operational observations:

• Retrieval from the structured relational map generally suffices for complex queries, minimizing token usage relative to vector-chunking RAG.
• Affective indexing is substantively valuable: queries regarding historical emotional states are frequent and efficiently resolved.
• Auditability transforms user trust: errors are diagnosable to transparent system actions, not opaque model failures.

The evidence is a single-user, single-corpus study. While not benchmarked against Mem0, Zep, or structured-memory ensembles, it offers practical longevity and sustained utility. Comparative evaluations are scheduled, emphasizing capacity-wise benchmarking in future work.

The Métacalque Principle and Overlay Layer

The paper introduces the concept of the métacalque: a principled, operational overlay architecture enabling layered, superimposed, and interval-bounded annotations across the corpus. Drawing from cartographic and literary theoretical traditions, the métacalque transforms descriptive layering into executable, actionable primitives for agentic retrieval.

Implemented overlays include affective, code, structural-anchor, concept, and insertion overlays; others, such as project-thread boundaries and genre clustering, are under development. Each overlay is expressed as a SQL-queried layer with soft boundaries and composable extents, shifting agent memory from hermeneutic reading to mechanistic retrieval control.

Implications for AI Agents and Knowledge Sovereignty

MOSS’s structure supports sovereign ownership and full portability, directly addressing the needs of regulated domains and researchers under ethics oversight. It enables memory architectures for organizations, supporting multi-instance deployment and domain-specific overlays. The inductive methodology ensures applicability from individual cognitive lives to complex enterprise settings.

The paper frames MOSS as an instantiation of exocortical extended memory [Clark & Chalmers, 1998], with practical, daily agentic interaction serving as empirical ground for extended-mind research.

Limitations and Future Directions

The deterministic retrieval execution confines non-determinism to agentic query formulation, recognized as a feature but not a complete solution. Construction costs of the inductive ontology for new corpora remain uncharacterized. Multimodal enrichment (audio, video, image) is ongoing, and future work includes multi-tenant isolation and advanced benchmark-driven evaluation.

Priority efforts include benchmarking against vectorial and structured-memory competitors, multi-researcher deployments under research-ethics frameworks, and comprehensive development of the métacalque overlay layer as a central architectural primitive.

Conclusion

MOSS exemplifies a rigorous inversion of mainstream RAG, centering auditable, sovereign, structurally unbounded, locally queryable memory for AI agents. By combining explicit relational structure, inductive ontology, agent-driven query formulation, and deterministic symbolic execution, MOSS enables AI agents to accompany individuals and organizations across years rather than sessions. This architecture addresses theoretical gaps and practical requirements in agentic memory, supporting transparent, scalable, and adaptive companion memory systems for autonomous agents (2607.04391).