Brain-Inspired Graph Memory Learning (BGML)

• BGML is a brain-inspired framework that utilizes graph structures to mimic key biological memory traits such as core–periphery organization and synaptic plasticity.
• It leverages modularity and decentralized updates, integrating mechanisms like logarithmic reinforcement and targeted pruning to support continual encoding and unlearning.
• Empirical results demonstrate BGML’s efficacy in enhancing memory retention, mitigating catastrophic forgetting, and optimizing agent strategies across diverse tasks.

Brain-Inspired Graph Memory Learning (BGML) encompasses a set of frameworks, models, and algorithms for memory-augmented graph-based learning, explicitly drawing from principles in cognitive neuroscience, neuroanatomy, and brain dynamics. BGML seeks to emulate hallmark traits of biological memory—core–periphery separation, hierarchical modularity, continual trace formation and erasure, local and decentralized plasticity, and synaptic consolidation—in dynamical graph structures or graph neural networks (GNNs). BGML methodologies thus aim for cognitive plausibility, long-horizon retention, efficient unlearning, and robust, interpretable memory representations.

1. Cognitive and Biological Motivations

BGML draws its theoretical foundation from the structure and function of biological memory systems. Core neurobiological phenomena inform its design:

• Core–Periphery Memory Architecture: Human declarative memory exhibits a sparse, highly consolidated "core" of identity-defining propositions and survival-relevant facts that are highly resistant to forgetting, juxtaposed against a periphery of less important, labile memories susceptible to decay and interference (Mollakazemiha et al., 2023).
• Multiple Interacting Memory Systems: Cognitive studies establish the presence of short-term working memory, rapid episodic encoding (hippocampal indexing), and slow cortical semantic memory acquisition, each with distinct timescales and retrieval mechanisms (Ma et al., 2022).
• Plasticity, Forgetting, and Stochasticity: Biological memory is inherently imperfect, supporting forgetting, interference, consolidation, and abrupt, stochastic changes reflecting emotion, attention, or character shifts (Mollakazemiha et al., 2023).
• Decentralization and Locality: Embodied in neuroscience via local synaptic updates, resource-limited microcircuits, and pervasive lack of a "global controller" (Wei et al., 2023, Wei et al., 2023).

BGML frameworks aim to preserve these biological characteristics within graph-based computational models, forming the conceptual bridge between cognitive neuroscience and AI memory systems.

2. Mathematical Formulations of BGML

BGML encompasses a spectrum of graph-based formal models capturing memory as modifiable graph structures:

a. Mass-based undirected memory graphs

• State: $G = (V,E)$, $V = \{1,\dots,n\}$ nodes as atomic propositions, $E$ undirected edges.
• Node attributes: Each $i$ has time-varying mass $M_i^t > 1$.
• Edge attributes: Each $\{i,j\}$ (if present) has weight $W_{ij}^t > 1$.
• Sequential update: Environmental input triggers edge or node addition; masses and weights evolve by logarithmic increments and stochastic log-Cauchy drift. Edge deletion ("pruning") occurs for $W_{ij}^t < \theta_w$. This yields consolidation in the core and decay in the periphery (Mollakazemiha et al., 2023).

b. Directed, resource-constrained graphs

• State: $G = (V,E)$, each node $v$ implements a microcircuit with $m$ inputs, $n$ outputs, and adaptive per-pair resistors $R_{x_i\to y_j}(t)$.
• Local update rule: Under conservation $\sum_{i,j} R_{x_i\to y_j}=R_c$,

$$\Delta R_{x_i\to y_j}(t) = \begin{cases} -R_{x_i\to y_j}(t)L_r, & I_{x_i\to y_j} > I_t \ +C/(mn-k), & \text{otherwise} \end{cases}$$

with $C$ the sum of decreased resources among "active" currents.

• Retrieval: As circuit simulation via Kirchhoff's laws, modes include upstream-only or cue-based "awakening." (Wei et al., 2023)

c. Autonomous active-directed graphs

• State: Nodes $x_i \in \{0,1\}$ with directed, local field-of-view; each node maintains an index table of activation input-output patterns.
• Algorithm: Stimulus propagation and consolidation rely exclusively on local fan-in/fan-out at each node, with entry merges controlled by capacity constraints. No global visibility or synchronization exists (Wei et al., 2023).

d. Hierarchical, modifiable GNN ensembles

• Partition: Initial graph is recursively partitioned into shards and subshards (e.g., via BEKM or BLPA). Feature graph construction encodes node-level information within grain-local GNNs.
• Update: As new nodes or forgetting requests arrive, only the locally relevant submodels are retrained or updated, following an Information Self-Assessment Ownership module to determine attachment and context (Miao et al., 2024).

3. Functional Mechanisms for Memory Operations

BGML supports a diverse array of biological and algorithmic memory operations:

Operation	BGML Mechanism	Reference
Remembering	Mass growth/logarithmic reinforcement	(Mollakazemiha et al., 2023)
Forgetting	Peripheral mass/weight decay, edge pruning	(Mollakazemiha et al., 2023)
Continual Encoding	Decentralized resistor update via inputs	(Wei et al., 2023)
Unlearning	Targeted scratch-retraining under FR requests	(Miao et al., 2024)
Incremental Learning	Feature similarity for context retrieval	(Miao et al., 2024)
Catastrophe Avoidance	Modularity/isolation of memory traces	(Miao et al., 2024)

In mass-based BGML, forgetting is driven by natural decay and pruning of low-weight edges, mimicking rapid periphery loss and stable core persistence. In locally adaptive, resource-limited networks, competition for synaptic "resources" enforces capacity and interference, revealing analogs of biological retroactive inhibition and rehearsal improvements (Wei et al., 2023). Autonomous, table-driven systems implement robust, combinatorial memory via WCC permutations, with empirical evidence for fault tolerance, parallelism, and resilience under massive damage or missing data (Wei et al., 2023).

4. BGML in Graph Neural Networks and Agent Systems

BGML mechanistically informs both theoretical GNN architectures and practical memory modules for autonomous agents:

• Working, Episodic, Semantic Memory Modules: Memory-augmented GNNs instantiate working memory via gated recurrence (e.g., GGNN), episodic memory via key–value or persistent state, and semantic memory with external centroid or virtual nodes, capturing hippocampal and cortical division of labor (Ma et al., 2022).
• Adaptive Meta-Cognitive Memory Graphs: LLM-based agent systems instantiate BGML via heterogeneous, multi-layered memory graphs ($Q$ task nodes, $T$ transition nodes, $M$ meta-cognition nodes) with reinforcement-optimized, dynamically weighted edges, driving strategic behavior via empirical utility signals (Xia et al., 11 Nov 2025).
• Oscillatory Graph Synchronization: HoloGraph and HoloBrain models bridge BGML with neurophysics, modeling each node as a neural oscillator steered by control vectors $y_i$ (encoded "memory"), enforcing synchronization patterns that prevent over-smoothing and enable long-range context propagation (Dan et al., 20 Jan 2026).

5. Empirical Results and Task Benchmarks

BGML frameworks have been quantitatively and qualitatively validated on a diversity of tasks:

• Classification and Continual Learning: On datasets such as Cora, CiteSeer, PubMed, and Coauthor-CS/Physics, BGML achieves micro-F1 scores exceeding 88–97%, outperforming MLP, ChebNet, GCN, and FGNs baselines. It demonstrates higher memory retention and lower catastrophic forgetting in unlearning and class-incremental protocols (Miao et al., 2024).
• Memory Trace and Retrieval: Decentralized directed-graph models and active-directed graphs yield capacity scaling that exceeds classical Hopfield networks for shallow depth, with robust retrieval fidelity ($\alpha\geq0.9$) in the presence of interference, fault, or partial cues (Wei et al., 2023, Wei et al., 2023).
• Agent Strategy Optimization: LLM agents with trainable graph memories show up to +25.8% zero-shot improvement (Qwen 3-4B, PopQA), +13.6% RL performance gains, and 2–4 pp ablation drops without memory updates, substantiating BGML's contribution to robust meta-cognitive control (Xia et al., 11 Nov 2025).
• Cognitive Connectomics: CogGNN fuses GNN-generated connectomic templates with ESN-based cognitive reservoirs, yielding CBTs (connectional brain templates) that are structurally sound, discriminatively powerful (66% for AD/LMCI, AUC=0.70), and superior in visual memory retention compared to non-BGML baselines (Soussia et al., 13 Sep 2025).

6. Limitations and Open Challenges

While BGML robustly models key aspects of biological memory, several open challenges remain:

• Theoretical Analysis: Formal characterization of expressivity and the effect of memory modules on GNN receptive fields, attention, and oversquashing remains outstanding (Ma et al., 2022).
• Parameter Selection and Stopping Criteria: Mass-based models require heuristic specification for initial mass/weight values, stochasticity controls, and pruning thresholds (Mollakazemiha et al., 2023).
• Scalability and Resource Constraints: Persistent and episodic memory systems risk unbounded growth; capacity management via modularity, resource conservation, and local plasticity remains an area for further investigation (Wei et al., 2023, Wei et al., 2023).
• Biological Mapping: Most models lack formal comparison with behavioral or neuroimaging data; extensions to batch/streaming, synaptic drift, and hierarchical brain modules remain largely unexplored.
• Empirical Benchmarks: Curated graph benchmarks targeting memory-specific capabilities (episodic/semantic distinction, selective forgetting) are required for fine-grained evaluation (Ma et al., 2022).

7. Extensions and Applications

BGML offers a flexible substrate for a range of current and future applications:

• Lifelong Learning and Unlearning: Efficient, localized adaptation to node/edge additions and deletions supports scalable, brain-like memory formation and erasure without global retraining (Miao et al., 2024).
• Neuromorphic Hardware and Simulations: The explicit mapping to local synaptic updates and resource limitation provides a blueprint for scalable hardware implementations and experimental investigation of trace theory (Wei et al., 2023).
• Interpretable Autonomous Agents: Explicit memory graphs with hierarchical abstraction and adaptive consolidation enable high-capacity, continually evolving, human-interpretable meta-strategies in complex decision-making (Xia et al., 11 Nov 2025).
• Connectomic and Cognitive Biomarker Discovery: Embedded cognitive and visual-memory tests within connectomics pipelines allow for population-level and individual-level mapping of function-structure relationships (Soussia et al., 13 Sep 2025).
• Graph Optimization and Reasoning: Oscillatory synchronization mechanisms support robust reasoning across dynamic or non-Euclidean graphs, improving generalization and combating numerical pathologies like over-smoothing (Dan et al., 20 Jan 2026).

In summary, Brain-Inspired Graph Memory Learning encapsulates a set of architectures, update rules, and design principles derived from neurobiological and cognitive insights, operationalized as modifiable graph memory systems. These frameworks combine modularity, stochasticity, local autonomy, and hierarchical abstraction to offer interpretable, robust, and scalable lifelong memory for graphs and graph-based learning.