Bibliography

Synix builds on several lines of prior work. This page maps the connections — where the project is borrowing ideas, where it is reframing them, and where it is making stronger claims that still need validation.

Cognitive architectures

CoALA (2024) — Modular cognitive architecture framework. Synix shares the modular framing and adds CDR constraints and multi-agent governance.

ACT-R / Soar — Classical architectures with memory subsystems. Informative for structural insights; Synix targets distributed multi-agent systems rather than single cognitive models.

Theoretical foundations

Hopfield Networks — Hopfield, 1982. Content-addressable memory via energy minimization in recurrent networks. Established that memory retrieval can be modeled as convergence to attractors in a dynamical system — a foundational result for understanding how associative memory works and degrades.

Modern Hopfield Networks — Ramsauer et al., 2020. Exponential storage capacity and connection to transformer attention. Shows that the attention mechanism in transformers is performing a form of associative memory retrieval, bridging classical memory models and modern architectures.

Nobel Prize in Physics 2024 — Awarded to Hopfield and Hinton for foundational discoveries enabling machine learning with artificial neural networks. Recognizes Hopfield networks as a key contribution to the field.

Complementary Learning Systems — McClelland et al., 1995. Hippocampal-neocortical consolidation theory. CLS identifies the division between fast learning and slow consolidation. CDR takes that as part of its motivation and adds a stronger argument for explicit renewal.

Cortical timescale hierarchy — Hasson et al. Different cortical regions process information at different temporal receptive fields. CDR uses a related intuition for partitioning by timescale.

Active Inference — Friston. Free energy minimization for adaptive systems. CDR shares the adaptation framing but focuses on memory structure and lifecycle rather than the inference mechanism itself.

Information-Theoretic Bounded Rationality — Genewein et al. Information-cost tradeoffs in decision-making. Provides one of the main theoretical precedents for treating memory and decision systems as resource-bounded.

DLM (March 2026) — Three-system autonomous learning with learnable consolidation schedules. DLM is especially relevant because it treats consolidation as a first-class problem; CDR pushes further toward explicit renewal.

Why AI Systems Don’t Learn and What to Do About It — Dupoux, LeCun & Malik, 2026. Proposes a cognitive science-inspired architecture with Systems I (fast perception), II (deliberate reasoning), and M (memory), controlled by meta-signals that switch between observation-based and behavior-based learning. Directly relevant to CDR’s argument that memory lifecycle needs explicit architectural support.

AGM Postulates & Iterated Belief Revision — Alchourrón, Gärdenförs & Makinson, 1985; Darwiche & Pearl. Formal theory for rational belief change through epistemic state revision.

Information theory & compression

MemFly (2026) — Formulates agent memory as Information Bottleneck optimization between relevant information and storage cost.

Rate-Distortion Framework for Summarization — Arda & Yener, 2025. Defines summarizer rate-distortion function as fundamental lower bound for compression.

Mutual Information Surprise (2025) — Redefines surprise as epistemic growth; triggers memory formation for high-impact observations.

Forgetful but Faithful (2025) — MaRS architecture with provenance-tracked memory nodes and six formal forgetting policies.

Agent memory systems

Generative Agents — Park et al., 2023. Introduces memory stream with reflection and three-factor retrieval scoring (recency, importance, relevance).

MemGPT / Letta — Packer et al., 2023. Agents managing their own memory via tools. Applies OS metaphor with tiered memory (context window as main memory, archival store as disk).

Hindsight — Latimer et al., 2025. Four epistemically-typed memory networks with parallel retrieval fused via RRF.

A-MEM — Xu et al., NeurIPS 2025. Zettelkasten-inspired autonomous note system with self-organizing memory graphs and significant multi-hop reasoning improvements.

Memory-R1 — Yan et al., 2025. Focuses on write-path optimization via RL to learn when and what to write.

Zep / Graphiti — Bi-temporal knowledge graph with contradiction detection. Strong production data model and one of the more serious references for contradiction handling.

Cog — Convention-based plain text memory where documents are the protocol and rules self-evolve. Closest existing system to Synix’s approach. Synixolis attempts to formalize the consistency model and multi-agent coordination that Cog leaves informal.

Mem0 — Extraction-based dual vector+graph memory. Good reference point for practical retrieval-oriented memory systems.

Memory operating systems

EverMemOS (2026) — Lifecycle system with episodic formation, semantic consolidation, and reconstructive recollection.

TiMem (2026) — Temporal Memory Tree with progressive consolidation based on age.

MemOS — MemTensor, 2025. Three-layer infrastructure-oriented memory OS with MemCube abstraction and lifecycle management.

MemoryGraft (2025) — Demonstrates memory security vulnerabilities and validates need for versioning and integrity checks.

Benchmarks

LongMemEval — Wu et al., ICLR 2025. Tests five memory abilities including extraction, reasoning, temporal tracking, and knowledge updates.

LoCoMo — Maharana et al., ACL 2024. Evaluates memory over naturalistic, multi-session conversations with realistic messy interactions.

Evo-Memory — DeepMind. Memory evolution benchmark, closest existing work to regression testing for memory systems.

LLMs Do Not Have Human-Like Working Memory — Huang et al., 2025. Demonstrates LLMs are stateless text processors requiring external memory infrastructure.

Emergence & adaptive dynamics

Emergence of Hierarchies in Multi-Agent Self-Organizing Systems (2025) — Demonstrates hierarchies emerge dynamically from joint objectives via gradient analysis.

Emergence of Hybrid Computational Dynamics Through RL (2025) — RL agents spontaneously develop hybrid attractor architectures for stable and flexible reasoning.

Active Inference for Multi-Agent Coordination (2025) — Framework balancing information gain against coordination costs in agent systems.

BRAIN: Bayesian Reasoning via Active Inference (2025) — Continuous Bayesian belief updating handling distribution shifts without retraining.

Distributed systems

Event Sourcing / CQRS — The append-only log and materialized views in Synixolis are direct applications of event sourcing.

ESAA — Event sourcing for autonomous agents. Synixolis extends this from task orchestration to memory governance.

Multi-Agent Memory from a Computer Architecture Perspective (March 2026) — Identifies the multi-agent memory consistency problem as structurally similar to CPU cache coherence.

Ray — Moritz et al., 2018. Distributed execution framework implementing heterogeneous actors at different cadences.

Matrix — Meta AI, 2025. Ray-native multi-agent framework achieving 2-15x throughput over centralized approaches.

A History of Erlang — Joe Armstrong, HOPL III, 2007. Design story of Erlang/OTP emphasizing process isolation and periodic rebuild over incremental repair.

Generational Garbage Collection — Lieberman & Hewitt 1983, Ungar 1984. Objects die young; partition by age for efficient collection. Direct analogue to CDR’s timescale partitioning.

The Actor Model — Hewitt et al., 1973. Concurrent computation via independent processes with message passing and no shared mutable state.

The Java Virtual Machine Specification — Oracle. The JVM’s memory model — heap partitioning, generational GC, class loading, and runtime data areas — is a direct precedent for CDR’s approach to memory lifecycle, region partitioning, and the distinction between hot and cold state.

Agent runtimes & infrastructure

Rivet agentOS — Rivet, 2026. Open-source agent operating system built on WebAssembly and V8 isolates. Lightweight runtime infrastructure for AI coding agents with ~6ms cold starts. Relevant as infrastructure-layer thinking about what agents need to run — Synixolis operates at the memory and state layer above this.

LangGraph, CrewAI, AG2/AutoGen, OpenAI Agents SDK — Production multi-agent frameworks with varying approaches to shared state. Relevant more as orchestration baselines than as direct answers to long-lived memory governance.

Protocols

MCP — Model Context Protocol. Universal tool transport. No memory semantics.

A2A — Agent-to-Agent protocol (Google). Agent discovery and task delegation. Complementary to Synixolis (A2A handles discovery; Synixolis handles shared state).