title: “SSM Novel Applications for Agent Memory”

Novel SSM Applications for Agent Memory: Deep Research

This document surveys creative, unexpected, and underexplored applications of state-space models (SSMs) across diverse fields, with specific attention to how each finding could be applied to Signet’s agent memory system. The standard applications (language modeling, recommendation, image classification) are covered in the companion literature review. This document focuses on the weird, the lateral, and the genuinely novel.

Companion documents:

• Literature Review — foundational papers
• Implementation Survey — production deployments, Rust ecosystem
• Integration Synthesis — how SSMs map to Signet’s pipeline

title: “SSM Novel Applications for Agent Memory”

1. Anomaly Detection: SSMs as Behavioral Drift Sensors

1.1 KambaAD: Kolmogorov-Arnold + Mamba for Time Series Anomalies

• Source: KambaAD (OpenReview 2025)
• Key Insight: Fuses Kolmogorov-Arnold Networks (KAN) with Mamba’s selective SSM. KAN enforces data consistency for rapid anomaly detection; Mamba captures anomalies caused by local variations; the internal selection mechanism handles distribution shifts without retraining.
• Signet Application: The significance gate today uses fixed thresholds (minTurns=4, noveltyThreshold=0.4). KambaAD’s architecture suggests a learned gate that adapts per-user: the SSM hidden state encodes the user’s normal interaction pattern, and a KAN-style consistency check flags sessions that deviate from this norm. When a developer suddenly starts asking about DevOps instead of frontend work, the anomaly score spikes, triggering more aggressive extraction. The selection mechanism handles the gradual drift case too — when a user’s focus shifts over weeks, the SSM adapts without triggering false alarms.

1.2 MambaITD: Insider Threat Detection from Behavioral Sequences

• Source: MambaITD (2025)
• Key Insight: Encodes heterogeneous enterprise logs into behavioral sequences using a Mapping ID combining action type, device context, and temporal segment. A Mamba encoder captures long-range dependencies in behavioral + interval sequences. Cross-modal fusion dynamically weights behavioral vs temporal information. Adaptive threshold optimization per-user by analyzing probability distributions. Achieved 91.31 F1 on CERT dataset, 1.36x faster than Transformer, 13% lower GPU memory.
• Signet Application: The behavioral Mapping ID concept maps directly to hook events. Each agent interaction can be encoded as (action_type={prompt|tool_call|file_edit|search}, context={project| session_type}, time_segment={morning|afternoon|sprint|idle}). The SSM then learns normal patterns per user. When the pattern deviates — a new project starts, the user’s rhythm changes, or a session type never seen before appears — the system can trigger proactive memory operations (broader search, new arc creation, identity refresh). The interval sequence modeling is particularly valuable: Signet could detect that a user who normally interacts every few minutes has gone silent for hours, suggesting a context switch.

1.3 MAAT: Mamba Adaptive Anomaly Transformer

• Source: MAAT (2025)
• Key Insight: Combines sparse attention for long-range dependency capture with a Mamba block for local variation sensitivity. An adaptive gating mechanism fuses both pathways. A skip connection and Gated Attention adaptively weight features from attention and SSM branches, improving anomaly localization.
• Signet Application: Contradiction detection. The current system uses 34 hardcoded antonym pairs. MAAT’s dual-pathway architecture suggests a contradiction detector where the attention pathway captures global semantic relationships (the same entity discussed in different ways across sessions) while the SSM pathway captures local temporal shifts (this week’s statement contradicts last week’s). The adaptive gating learns which pathway matters for which entity domain.

title: “SSM Novel Applications for Agent Memory”

2. Code Understanding: SSMs as Session Transcript Parsers

2.1 CodeSSM: What State Space Models Learn About Code

• Source: CodeSSM (2025)
• Key Insight: SSMs outperform comparable Transformers at capturing code properties during pretraining — superior syntactic capture and semantic understanding in deeper layers. Introduces SSM-Interpret, a frequency-domain analysis showing SSM convolution kernels function as spectral filters (low-pass for long-range, high-pass for short-range). Critical finding: fine-tuning causes CodeSSM to lose syntactic information through a “spectral shift” emphasizing short-range dependencies. Complementary kernel behavior (forward high-pass + backward low-pass) captures the most code properties.
• Signet Application: Agent session transcripts are saturated with code. Today’s extraction pipeline treats code blocks as opaque text. An SSM pre-trained on code could: (1) identify which code entities in a transcript are structurally significant (function signatures vs debug print statements), (2) extract semantic relationships between code entities (dependency graphs, call hierarchies) that inform the knowledge graph, (3) detect when code discussion reveals architecture decisions worth remembering vs routine implementation details. The spectral filter insight is particularly relevant: different SSM layers naturally separate high-frequency detail (variable names, syntax) from low-frequency structure (architectural patterns, design decisions) — exactly the distinction the extraction pipeline needs.
• Warning: The “spectral shift” during fine-tuning means care is needed when adapting a code-pretrained SSM to memory-specific tasks. The model may lose its long-range code understanding if fine-tuned aggressively on short memory classification tasks.

title: “SSM Novel Applications for Agent Memory”

3. Planning and Decision-Making: SSMs as Anticipatory Engines

3.1 Decision Mamba: Multi-Grained SSM for Offline RL

• Source: Decision Mamba (NeurIPS 2024)
• Key Insight: Dual-branch architecture with coarse-grained SSM (InterS3M) for global sequential dependencies and fine-grained SSM (IntraS3M) for local state-action-return triplets. Progressive Self-Evolution Regularization (PSER) blends prior predictions with ground truth: a_refined = (1-beta)a_true + betaa_prior. Achieves 83.2% normalized score on D4RL Gym-MuJoCo, surpassing Decision Transformer (75.8%).
• Signet Application: The dual-granularity architecture maps to session-level vs turn-level memory prediction. The coarse SSM models the arc of an entire session (what broad topic is being explored), while the fine SSM models individual turns (what specific memory will be needed for the next prompt). PSER is directly applicable to Signet’s prediction refinement — blend the predictor’s previous guess with ground truth feedback to progressively improve without overfitting to noisy training signals. The key insight: offline RL trajectories have hierarchical structure. So do agent sessions.

3.2 Drama: SSMs as World Models

• Source: Drama (ICLR 2025)
• Key Insight: First Mamba-based world model for model-based RL. Only 7M parameters vs DreamerV3’s 200M (28x reduction). Discrete VAE encodes observations into latent embeddings; Mamba-2 processes concatenated latent states and actions to predict futures. Dynamic Frequency-based Sampling (DFS) favors transitions the world model has learned sufficiently. Trains on standard laptop hardware. Mamba’s input-dependent design automatically resets hidden states at episode boundaries.
• Signet Application: Drama proves SSMs can serve as world models with extreme parameter efficiency. For Signet, a 7M-parameter “session world model” could predict what the user will do next — not just what memory to retrieve, but what the entire next session fragment will look like. The DFS insight is directly applicable: prioritize training on session transitions the model has already learned, avoiding wasted compute on rare edge cases. The automatic episode boundary detection maps to session start/end detection — no manual reset logic needed in the session tracker.

3.3 Hierarchical Decision Mamba (HDM)

• Source: HDM (2025)
• Key Insight: Stacked hierarchy with meta-Mamba (high-level sub-goal planning) and control-Mamba (low-level action execution). The meta level sets objectives; the control level executes within those constraints.
• Signet Application: Maps directly to Signet’s session-arc-epoch hierarchy. A meta-SSM operates at the arc level (what broad objective is the user pursuing over multiple sessions?) while a control-SSM operates at the session level (what specific memories does this session need?). This replaces the hardcoded arc threshold (8 sessions) and epoch threshold (4 arcs) with learned boundaries.

title: “SSM Novel Applications for Agent Memory”

4. Multi-Agent Coordination: SSMs for Shared Identity

4.1 Multi-Agent Memory as Computer Architecture

• Source: Multi-Agent Memory (2025)
• Key Insight: Proposes three-layer memory hierarchy mirroring classical computer architecture: I/O layer (ingestion), Cache layer (fast limited-capacity for immediate reasoning — KV caches, embeddings), Memory layer (persistent retrieval — vector DB, graph DB). Identifies two critical unsolved problems: (1) no principled protocol for sharing cached artifacts across agents (analogous to multiprocessor cache coherence), (2) under-specified access control for multi-agent memory (scope: documents, chunks, records, or trace segments?). The most pressing challenge: memory consistency models specifying which updates become visible to which agents and in what order.
• Signet Application: This is Signet’s multi-agent future described in architectural terms. The paper validates Signet’s existing hierarchy (working memory cache + SQLite persistent store + graph DB) but identifies the exact gap Signet will face: when multiple agents share an identity, how do you handle concurrent memory writes? The SSM angle: each agent maintains a local SSM hidden state representing its session trajectory. Synchronization happens by merging hidden states (averaging, weighted combination, or concatenation + reduction) rather than merging raw memories. This is dramatically more efficient than reconciling individual memory records.

4.2 Communicating Plans via World Models

• Source: Plan Communication (2025)
• Key Insight: Instead of communicating raw observations between agents, use a compact world model to simulate future states. A Message Generation Network compresses plans into compact messages. Engineered world model approach shows superior performance, sample efficiency, and scalability vs emergent communication.
• Signet Application: When multiple Signet agents share identity, they don’t need to share entire session transcripts or memory deltas. Instead, each agent’s SSM world model generates a compact “session plan summary” — what was accomplished, what changed, what the user’s trajectory looked like — and sends that compressed state vector to other agents. This is orders of magnitude more bandwidth-efficient than the current approach of syncing raw memories.

title: “SSM Novel Applications for Agent Memory”

5. NLU Tasks: SSMs for Significance Gating and Classification

5.1 Mamba for Long Text Classification

• Source: [Mamba Long Text](https://ieeexplore.ieee.org/document/ 10819509/) (2025)
• Key Insight: Mamba-130m achieves strong balance between efficiency and performance for long text classification, with input-dependent parameters enabling content-aware information propagation. Linear complexity makes it suitable for classifying very long documents.
• Signet Application: Session transcripts are long documents (frequently 20k+ characters). Classification tasks include: significance assessment (worth extracting?), topic detection (which arc does this session belong to?), sensitivity classification (does this contain secrets?). A small Mamba classifier could replace the current regex-based significance gate with a learned gate that considers the full session content, not just turn count and entity overlap.

5.2 RankMamba and SSMs as Text Rerankers

• Source: RankMamba (2024); SSM Rerankers (2024)
• Key Insight: Mamba achieves competitive document ranking with transformers at similar model sizes. The Mamba Retriever excels in inference speed with linear time scaling, particularly suited for long-text retrieval. Mamba-2 outperforms Mamba-1 in both performance and efficiency. Current practical limitation: training throughput lags behind flash attention-optimized transformers.
• Signet Application: Memory retrieval reranking. The current pipeline does alpha-blended BM25 + vector (0.7/0.3) with a top-5 cutoff. A Mamba reranker could take the top-20 candidate memories and rerank them using the full session context as a query, with linear time complexity enabling real-time reranking even during streaming. The reranker maintains a hidden state across the session, so later reranking queries benefit from the accumulated session context — something the current stateless per-turn approach cannot do.

5.3 ss-Mamba: Semantic-Spline Selective SSM

• Source: ss-Mamba (2025)
• Key Insight: Integrates semantic-aware embeddings within a selective state-space framework. Enhances forecasting by making the SSM aware of semantic context, not just numerical patterns.
• Signet Application: Memory embeddings today are static Nomic vectors computed once at storage time. An ss-Mamba-style approach could produce context-dependent embeddings that change meaning based on the current session’s semantic context. “Python” in the context of a snake discussion vs a programming discussion would produce different effective embeddings, improving retrieval precision without re-embedding the entire memory store.

title: “SSM Novel Applications for Agent Memory”

6. Event Stream Processing: SSMs for Hook Event Pipelines

6.1 SSMs for Complex Event Detection in CPS-IoT

• Source: CPS-IoT Foundation (2025)
• Key Insight: Mamba outperforms LSTM, TCN, causal Transformers, and neurosymbolic FSMs for online complex event detection from sensor traces. A 12-block Mamba (1.8M params, hidden=128, state=64) achieves F1~0.92 on standard data and 0.75 on 30-minute out-of-distribution sequences (trained on 5-minute sequences only). Key: Mamba generalizes to sequences 6x longer than training data.
• Signet Application: The daemon’s hook pipeline is a complex event stream: session_start, user_prompt, tool_call, file_edit, session_end, interleaved across multiple concurrent sessions. Today each hook is processed independently. An SSM event detector could recognize complex event patterns spanning multiple hooks: “the user started a session, asked about X, edited file Y, then asked about Z” is a pattern that should trigger specific memory operations. The 6x generalization is critical — the model can learn from short sessions and generalize to marathon sessions.

6.2 Compute-in-Memory SSM Implementation

• Source: [CiM SSM](https://www.nature.com/articles/s41467-025- 68227-w) (Nature Communications 2025)
• Key Insight: Implements SSMs in energy-efficient compute-in-memory hardware for event-driven processing. Event-SSM requires 1.68 GFLOPs for DVS128 Gesture, vs ResNet-18’s 104.28 GFLOPs (62x reduction). Processing is truly event-driven: computation only occurs when state-changing events arrive, with idle periods consuming near-zero power.
• Signet Application: Validates the thesis that SSMs can be event-driven rather than clock-driven. Signet’s hook pipeline is inherently event-driven — hooks fire when the user does something, not on a fixed schedule. An SSM that only computes when hooks arrive (rather than continuously sampling) would be maximally efficient. This architecture means the SSM sidecar could sit at near-zero CPU when the user is idle, waking only when hooks fire.

title: “SSM Novel Applications for Agent Memory”

7. Hierarchical Reasoning: Multi-Scale Temporal Memory

7.1 MS-SSM: Multi-Scale State Space Model

• Source: MS-SSM (ICLR 2025)
• Key Insight: A multi-resolution framework that represents sequence dynamics across multiple levels of detail simultaneously, capturing both fine-grained high-frequency patterns and coarse low-frequency trends.
• Signet Application: Directly maps to Signet’s temporal hierarchy: turn-level (what to retrieve now), session-level (arc tracking), arc-level (project tracking), epoch-level (identity evolution). Today these are handled by separate hardcoded systems (session tracker, arc summarizer, epoch condensation). A multi-scale SSM processes all four resolutions simultaneously in a single model, with learned interactions between scales. When a turn-level anomaly (unusual query) propagates up to trigger an arc-level state change (new project detected), this happens through learned dynamics, not hardcoded thresholds.

7.2 Hierarchical Spatio-Temporal SSM for fMRI

• Source: [HSTSSM](https://link.springer.com/chapter/10.1007/978-3- 031-90252-9_6) (2025)
• Key Insight: Processes spatial and temporal information separately via hierarchical Mamba-based encoders, then fuses them. Discovers neurological biomarkers by decomposing spatio-temporal dynamics.
• Signet Application: The spatial/temporal decomposition applies to memory in a non-obvious way. “Spatial” in Signet’s context is the knowledge graph topology (which entities are connected, how). “Temporal” is the sequence of interactions. An SSM that processes graph structure and temporal dynamics separately, then fuses them, could answer questions the current system cannot: “Given the entity graph neighborhood of X AND the recent temporal trajectory of the session, what memory is most relevant?” — combining structural and temporal reasoning in a principled way rather than ad-hoc alpha blending.

title: “SSM Novel Applications for Agent Memory”

8. Causal Inference: SSMs for Understanding “Why”

8.1 CausalMamba: Causal Discovery from Temporal Sequences

• Source: CausalMamba (2025)
• Key Insight: Unifies Mamba sequence encoding, GCN graph encoding, and differentiable causal discovery (NOTEARS). Mamba encoder (2 layers, 128-dim hidden) captures sequential content dependencies; GCN encodes structural relationships; NOTEARS-based learner discovers weighted adjacency matrices representing causal relationships. Feature fusion: H = H_seq + 0.3*H_graph. L1 sparsity + acyclicity constraint ensures interpretable DAGs. Intervention simulation validates causal structure.
• Signet Application: The most creative application in this document. CausalMamba’s architecture could discover WHY certain memories become relevant. Instead of just tracking “memory X was retrieved when user asked Y,” a CausalMamba-style model discovers causal chains: “because the user changed to project Z, memory X about framework W became relevant, which caused retrieval of memory V about configuration.” This transforms the memory system from reactive (retrieve what matches) to explanatory (understand why things match). The causal DAG could also inform memory retention: memories that are causal ancestors of frequently-retrieved memories should decay slower, even if they themselves are rarely retrieved directly.

8.2 State-Aware Causal Inference

• Source: [Observational Causality](https://www.nature.com/articles/ s42005-025-02447-w) (Communications Physics 2025)
• Key Insight: Quantifies causality as information gain about future states. Can characterize causal influence as a function of system state and distinguish between redundant and synergistic interactions.
• Signet Application: Information-gain-based causality maps to memory importance. A memory’s importance is not just its similarity to the current query — it is the information gain it provides about the user’s future trajectory. A memory that, when retrieved, changes the prediction of what the user will do next is causally important. A memory that, when retrieved, doesn’t change the prediction is redundant and can decay faster.

title: “SSM Novel Applications for Agent Memory”

9. SSMs as Embedding Engines: Beyond Static Vectors

9.1 The Hybrid Attention-SSM Complementarity Finding

• Source: [Hybrid Architecture Survey](https://www.askaibrain.com/en/ posts/end-of-transformers-hybrids-attention-state-space-2025) (2025); [Jamba](https://proceedings.iclr.cc/paper_files/paper/2025/file/ a9ed43fa31dc8b4a7d7a673d713dcb5f-Paper-Conference.pdf) (ICLR 2025)
• Key Insight: A 2025 ablation study on Jamba found that removing attention layers causes retrieval accuracy to drop to 0%. SSM layers alone contribute nothing to associative recall. Pure SSMs fail on copy and recall tasks despite excelling at long sequences. Hybrids use Mamba for bulk sequence processing (7 of 8 layers) and attention for precision recall (1 of 8 layers).
• Signet Application: This is a critical design constraint. SSMs alone cannot replace the vector similarity search that Signet relies on for memory retrieval. The architecture MUST be hybrid: SSM for temporal state tracking, sequence compression, and pattern detection; attention (or vector similarity) for precise associative recall. The implication: the SSM does not replace Nomic embeddings. It augments them with temporal context. The SSM hidden state conditions the retrieval query, but the retrieval itself uses attention-like mechanisms (cosine similarity, cross-attention).

9.2 GrassNet: SSMs for Graph Spectral Filtering

• Source: GrassNet (2025, Pattern Recognition 2026)
• Key Insight: First use of SSMs to design GNN spectral filters. SSMs model correlations of graph signals at different frequencies, deriving unique rectifications for each frequency in the graph spectrum. Overcomes polynomial filter limitations (low-order truncation, identical treatment of numerically close frequencies). Superior performance on 9 benchmarks.
• Signet Application: The knowledge graph is a graph. GrassNet suggests using an SSM to learn spectral filters over the entity graph, producing entity embeddings that capture graph-structural information at multiple frequency scales. Low-frequency filters capture global community structure (which entities cluster together), while high- frequency filters capture local neighborhood variation (which entity is anomalous in its neighborhood). This replaces the current static graph traversal with learned graph reasoning that adapts to the graph’s spectral properties.

9.3 DyGMamba: Dynamic Graph Embeddings

• Source: DyGMamba (2025); DyG-Mamba (NeurIPS 2025)
• Key Insight: Two-level SSM for continuous-time dynamic graphs. Node-level SSM encodes historical interactions per entity; time-level SSM exploits temporal patterns to dynamically select critical information from interaction history. DyG-Mamba treats irregular timespans between events as control signals, allowing the model to dynamically adjust forgetting based on inter-event intervals.
• Signet Application: The knowledge graph is not static — entities gain and lose relevance over time, relationships form and dissolve. DyGMamba’s architecture maps to: (1) per-entity SSM that tracks each entity’s interaction history (when was it mentioned, in what context, how has its usage evolved), (2) a global temporal SSM that identifies which entities are currently “hot” based on cross-entity temporal patterns. The inter-event interval as control signal is particularly elegant: a long gap between entity mentions should increase forgetting (the entity is fading from relevance), while a burst of mentions should strengthen retention. This replaces the hardcoded 0.95^ageDays with learned, entity-specific decay dynamics.

title: “SSM Novel Applications for Agent Memory”

10. Edge Hardware: SSMs on User Machines Without GPUs

10.1 BrainChip TENN: Temporal Event-Based Neural Networks

• Source: [BrainChip TENN](https://brainchip.com/temporal-event- based-neural-networks-a-new-approach-to-temporal-processing/) (2025); Akida Cloud (2025)
• Key Insight: TENNs build on SSM architecture with event-driven processing — computation only happens when state-changing events arrive, skipping periods of no change. Combines spatial and temporal convolutions for fewer parameters and MACs per inference than traditional networks. Operates in temporal convolution mode (training) or recurrent mode (inference). Akida Pico FPGA Cloud launched Feb 2026, AKD2500 custom silicon targeting Q3 2026 (TSMC 12nm). Power consumption in tens of milliwatts for video object detection.
• Signet Application: TENN’s event-driven processing is the ideal architecture for Signet’s sidecar. The SSM should not run continuously — it should activate only when hooks fire (agent interaction events). Between events, power consumption is near-zero. The TENN approach of dual-mode operation (temporal convolution for training parallelism, recurrent for inference efficiency) maps to Signet’s needs: train the model offline using accumulated session data (parallel), then deploy as a recurrent model that processes one hook at a time (sequential). The Akida FPGA Cloud could enable Signet to train models remotely on neuromorphic hardware and download the trained model for local recurrent inference on CPU.

10.2 Mamba-X: End-to-End SSM Accelerator for Edge

• Source: Mamba-X (2025)
• Key Insight: Hybrid H2 quantization tailored to SSM data distributions, quantizing weights and activations to INT8. Addresses memory constraints of edge devices specifically.
• Signet Application: INT8 quantization of the SSM sidecar would reduce memory footprint from ~40MB (FP32, 10M params) to ~10MB (INT8). Combined with Mamba’s already-small parameter counts (Drama achieves world modeling at 7M params), a quantized Signet SSM could fit in 7-10MB of RAM and run on any modern CPU without GPU.

10.3 FastMamba: Hadamard-Quantized SSMs

• Source: [FastMamba](https://openreview.net/pdf/bd4cfd9e6528985d fbf397995a38e896464faa3b.pdf) (2025)
• Key Insight: 8-bit quantization via Hadamard transforms for outlier mitigation. 68.8x speedup over CPU, 8.9x over GPU for Mamba2-130M. 6x higher energy efficiency than RTX 3090. Power-of-two scaling for SSM and convolution blocks.
• Signet Application: 68.8x CPU speedup means a Mamba2-class model that takes 1 second per inference on baseline CPU would take 14.5ms with FastMamba quantization. For Signet’s per-hook processing, 14.5ms is well within the budget (hooks currently allow up to 500ms). The energy efficiency matters for laptop users — the SSM sidecar should not drain battery.

10.4 Fully Quantized Mamba in 1.58 Bits

• Source: [1.58-bit Mamba](https://aclanthology.org/2025.coling- main.316.pdf) (COLING 2025)
• Key Insight: Extreme quantization to ternary weights (-1, 0, 1) while maintaining usable performance. Replaces all multiplications with additions/subtractions.
• Signet Application: A 1.58-bit SSM sidecar at 10M parameters would occupy ~2MB. This is small enough to ship as a bundled binary inside the signetai npm package. No separate download, no model registry, no GPU detection. The model literally fits inside the CLI bundle. Performance would degrade from full precision, but for tasks like significance gating and importance decay (where precision matters less than speed), ternary weights may be sufficient.

title: “SSM Novel Applications for Agent Memory”

11. Continual Learning: SSMs That Never Forget

11.1 Mamba-CL: Null-Space Projection for Continual Learning

• Source: Mamba-CL (2024)
• Key Insight: Updates SSM parameters only in directions orthogonal to the feature subspace of previous tasks. Theoretically guarantees that new learning never interferes with old learning. Derives constraints on four critical time-invariant parameters within Mamba’s architecture. Superior results on 4 class-incremental benchmarks.
• Signet Application: The SSM sidecar will learn from the user’s interaction patterns. But users change over time — new projects, new tools, new coding styles. Mamba-CL’s orthogonal projection ensures that learning new patterns (user adopts Rust after years of TypeScript) doesn’t destroy the model’s knowledge of old patterns (TypeScript knowledge stays intact for when the user revisits those projects). This is exactly the continual learning guarantee Signet needs: the predictor gets better over time without catastrophic forgetting.

11.2 MemMamba: Solving SSM Long-Range Forgetting

• Source: MemMamba (2025)
• Key Insight: Proves mathematically that SSM long-range memory decays exponentially. Introduces state summarization (inspired by human document summarization) + dual attention (cross-layer and cross-token) to combat decay while preserving linear complexity. 48% speedup in inference. Breaks the complexity-memory trade-off.
• Signet Application: The exponential decay proof is both a warning and an opportunity. Warning: a naive SSM WILL forget early session content by the end of a long session. Opportunity: MemMamba’s state summarization mechanism mirrors Signet’s existing session summary system. The SSM could produce compressed “memory checkpoints” at intervals during long sessions, preventing information loss. The cross-layer attention enables memories at different abstraction levels (facts vs entities vs arcs) to interact, which the current pipeline handles through explicit joins.

11.3 MambaCL: Meta-Learning for Continual Learning

• Source: MambaCL (2024)
• Key Insight: Treats continual learning as next-token prediction. Introduces selectivity regularization bridging SSM selective parameters with Transformer attention weights via the duality between SSMs and linear attention. During training, ground-truth associations guide Mamba’s selective behavior via KL divergence loss. Key finding: Mamba matches or exceeds Transformer performance with fewer parameters, with particular strength on fine-grained tasks and length generalization (5x longer sequences than training).
• Signet Application: The selectivity regularization concept is directly transferable. During training, Signet knows which memories were actually useful (via the feedback signal). This ground-truth association can guide the SSM’s selective gating — when the SSM sees a session that resembles one where memory X was useful, it should increase its gate for features related to X. The 5x length generalization means a model trained on 10-turn sessions can work on 50-turn sessions.

11.4 Mamba-FSCIL: Few-Shot Class-Incremental Learning

• Source: Mamba-FSCIL (2024)
• Key Insight: Dual selective SSM projector with frozen base branch
  • trainable incremental branch. Class-sensitive selective scan with suppression loss (keep old patterns quiet) and separation loss (make new patterns distinct). Reduces performance drop on base classes while learning new classes from minimal examples.
• Signet Application: When a user starts a new project, the SSM needs to learn new patterns from very few sessions (few-shot). Mamba- FSCIL’s dual-branch approach keeps the base model frozen (general interaction patterns) while training a lightweight branch on the new project’s patterns. The suppression/separation losses ensure that recognizing “user is working on Rust project” doesn’t interfere with the model’s existing ability to recognize “user is working on TypeScript project.”

title: “SSM Novel Applications for Agent Memory”

12. Information-Theoretic Foundations: The Theory of SSM Memory

12.1 Mathematical Formalism for SSM Memory Compression

• Source: [Memory Compression Formalism](https://arxiv.org/html/ 2410.03158v1) (2024)
• Key Insight: Applies rate-distortion theory and information bottleneck to SSM gating. Key theorem: if gating ensures mutual information between compressed state and input exceeds I_min, then distortion is bounded by D_max(I_min). Selective gating achieves 30-60% reduction in hidden state utilization for long sequences without proportional performance loss. Fano’s inequality provides lower bounds on prediction error from compressed states. The gating function G(x_t) acts as a dynamic information filter: values near 0 retain previous state (persistence), values near 1 incorporate new input (flow).
• Signet Application: This formalizes what Signet’s memory system does intuitively. Every memory decision is an information compression decision: what to keep, what to discard, how much detail to retain. The rate-distortion framework provides principled bounds: given a target memory budget (say, 500 memories), what is the minimum information loss achievable? The gating function’s persistence/flow duality maps exactly to memory retention: high persistence = long-term memory (rarely updated facts), high flow = working memory (frequently revised state).

12.2 MPS-SSM: Minimal Predictive Sufficiency

• Source: MPS-SSM (2025)
• Key Insight: Seeks maximal compression of the past under the strict constraint that no predictive capability is lost. Unlike information bottleneck (which trades off compression vs prediction), MPS achieves BOTH maximal compression AND full predictive power by filtering exclusively non-predictive information. Provably converges to minimal sufficient statistics. Invariant to non-causal perturbations (noise robustness). Sweet spot: clean data needs modest regularization; noisy/complex data needs stronger regularization.
• Signet Application: This is the theoretical foundation for the entire memory system. Signet should not store everything and should not discard randomly. It should store the minimal set of memories that preserves full predictive power over the user’s future needs. MPS-SSM provides the mathematical framework: for each memory, ask “does this memory improve prediction of future retrieval queries?” If not, it can be discarded without loss. If yes, it must be retained. The regularization sweet spot insight means Signet should be more aggressive about compression in clean, predictable interaction patterns, and more conservative in novel/complex scenarios.

title: “SSM Novel Applications for Agent Memory”

13. Bonus: Forgetting Curves and Bioinspired Memory Dynamics

13.1 Human-Like Forgetting Curves in Neural Networks

• Source: [Forgetting Curves in DNNs](https://arxiv.org/html/ 2506.12034v2) (2025)
• Key Insight: MLPs exhibit Ebbinghaus-like memory decay: sharp initial loss followed by stabilization. Recall probability decays as power law. When recall drops to 80% of baseline, “review sessions” restore it — with progressively lengthening intervals (4, 10, 31 epochs), mirroring biological spaced repetition. Repeated reinforcement increases peak recall beyond pre-review levels.
• Signet Application: Validates that the SSM sidecar’s internal memory decay will follow biological patterns, not catastrophic forgetting. The spaced repetition finding is directly applicable to memory maintenance: instead of the current fixed 30-day retention window, schedule memory “review” based on the SSM’s internal recall probability. Memories whose SSM representation is fading get re-presented during training at progressively longer intervals. This is the neural implementation of VISION.md’s “learns what to remember.”

13.2 FOREVER: Forgetting Curve-Inspired Memory Replay

• Source: FOREVER (2026)
• Key Insight: Measures model evolution through parameter-space distance (L2 norm of parameter changes) rather than step counts. Calibrates “model days” from early training dynamics. Schedules replay at Ebbinghaus-curve intervals (1, 2, 4, 7, 15 model-days). Intensity-aware regularization adjusts replay strength based on current learning rate instability. Consistent 1.2% overall improvement, 0.9% backward transfer improvement across 0.6B-13B parameter scales.
• Signet Application: FOREVER’s “model-centric time” concept solves a real problem. The SSM sidecar’s training schedule should not be clock-based (retrain every N hours) but model-evolution-based (retrain when the model has drifted sufficiently from its last checkpoint). During active sessions with many new memories, the model evolves rapidly and needs frequent replay. During quiet periods, the model barely changes and replay is wasteful. The intensity-aware regularization prevents the model from over-adapting to a single intense session while remaining plastic for gradual evolution.

title: “SSM Novel Applications for Agent Memory”

14. Bonus: Multimodal Fusion and Cross-Modal Memory

14.1 Mixture-of-Mamba: Modality-Aware Sparsity

• Source: [Mixture-of-Mamba](https://openreview.net/forum?id= Valt8gMdfl) (2025)
• Key Insight: Modality-specific parameterization of the Mamba block enables efficient multi-modal processing. Matches image loss at 35% of training FLOPs; matches speech loss at 25% of FLOPs. Each modality gets its own sparse pathway through the SSM.
• Signet Application: Agent interactions are multimodal: text conversations, code blocks, file paths, URLs, tool outputs, error messages, screenshots (via browser extensions). Mixture-of-Mamba suggests that the SSM should have modality-specific pathways for different memory types. A code memory should activate different SSM parameters than a conversational memory. This reduces interference between modalities and improves efficiency: code-heavy sessions route through code-specific parameters, while planning sessions route through planning-specific parameters.

14.2 AlignMamba: Cross-Modal Alignment in SSMs

• Source: [AlignMamba](https://openaccess.thecvf.com/content/ CVPR2025/papers/) (CVPR 2025)
• Key Insight: Injects cross-modal alignment signals — both token-level (local) and distributional (global) — before the Mamba backbone processes the fused representation. Maintains linear computational complexity.
• Signet Application: When a memory contains both text and code (as most extracted memories do), AlignMamba’s approach suggests aligning the text and code representations before feeding them to the SSM. This ensures the SSM sees a coherent multimodal representation rather than a concatenated mess. The distributional alignment is particularly relevant for detecting when text descriptions of code diverge from the actual code behavior (a form of contradiction detection).

title: “SSM Novel Applications for Agent Memory”

15. Bonus: Brain Activity Encoding and Cognitive State

15.1 Brain-Mamba: Neural Signal Encoding via SSMs

• Source: [Brain-Mamba](https://proceedings.mlr.press/v248/ behrouz24a.html) (2024)
• Key Insight: Attention-free, scalable framework using SSMs to encode brain activity across multiple neuroimaging modalities. Two modules: MLP for multi-channel integration + S4 for temporal encoding; GNN for learning brain region interdependencies. Outperforms all baselines across 7 datasets and 3 modalities.
• Signet Application: The “brain region interdependency” concept maps to entity interdependency in the knowledge graph. Brain-Mamba’s architecture of encoding temporal dynamics per-channel (per-entity) and then learning cross-channel structure (entity relationships) is exactly what a graph-aware SSM for Signet’s knowledge graph should look like. Each entity gets its own temporal encoding, and a GNN layer learns which entities are interdependent — replacing the current static graph traversal with dynamic, learned entity relationship reasoning.

15.2 EEGMamba: Foundation Models for Neural Signals

• Source: [EEGMamba](https://www.sciencedirect.com/science/article/ abs/pii/S0893608025006963) (Neural Networks 2025)
• Key Insight: Foundation model for EEG using Mamba encoder, achieving SOTA across 6 downstream tasks. Demonstrates that SSMs can serve as general-purpose encoders for high-dimensional temporal signals, transferring representations across vastly different tasks.
• Signet Application: If EEGMamba can encode brain signals across 6 different tasks with a single pretrained model, a similar SSM foundation model could encode agent interaction patterns across different pipeline tasks (significance gating, extraction quality prediction, retention scoring, contradiction detection) with shared representations. Pre-train once on session data, fine-tune for each pipeline stage.

title: “SSM Novel Applications for Agent Memory”

16. Bonus: Test-Time Adaptation and Online Learning

16.1 Test-Time Training for Continuous Adaptation

• Source: TTT (2025)
• Key Insight: Equips sequence models with hidden states that function as lightweight parametric models updated online during inference. Contextual information compressed into state matrix S; the product Sq generates output. TTT-E2E achieves constant inference latency regardless of context length (2.7x faster than full attention at 128K context).
• Signet Application: The SSM sidecar could update its parameters during inference — not just during training. As each hook fires, the model not only produces predictions but also learns from the result. If the model predicted that memory X would be relevant but the user ignored it, the model adjusts in real-time. This is the ultimate expression of “gets sharper the longer you use it” from VISION.md — the model improves within a single session, not just between training runs.

title: “SSM Novel Applications for Agent Memory”

Synthesis: The Creative Map

The 10 research areas above, plus 6 bonus areas, converge on a coherent architecture. Here is how they compose:

What SSMs Replace (learned dynamics replacing heuristics)

What SSMs Add (capabilities that don’t exist today)

Design Constraints Discovered

1. SSMs cannot do retrieval alone (Sec 9.1). The architecture MUST be hybrid: SSM for temporal state + attention/similarity for recall.
2. SSMs forget exponentially (Sec 11.2). Long sessions need state summarization checkpoints.
3. Fine-tuning causes spectral shift (Sec 2.1). Code understanding degrades if fine-tuned aggressively on non-code tasks.
4. Training throughput lags flash attention (Sec 5.2). SSMs are faster at inference but slower to train than optimized transformers.
5. Quantization to INT8 is practical (Sec 10.2-10.3). 1.58-bit is achievable for coarse tasks.

The Minimal Viable SSM Stack for Signet

Based on the research above, the smallest useful SSM integration would be:

1. Event-driven session state SSM (~2M params): Processes hook events, maintains session trajectory, provides temporal context for retrieval queries. Based on Drama’s architecture (Sec 3.2) with event-driven activation (Sec 6.2).

2. Significance gate SSM (~1M params): Replaces regex-based significance assessment with a learned gate trained on extraction outcomes. Based on KambaAD (Sec 1.1) for anomaly-aware gating.

3. Importance decay SSM (~2M params): Per-entity learned decay replacing 0.95^ageDays. Based on DyGMamba (Sec 9.3) for dynamic graph-temporal modeling.

Total: ~5M parameters, ~10MB at INT8, ~2MB at 1.58-bit. Inference: <15ms per hook on CPU (Sec 10.3). Architecture: Rust sidecar, same as current predictor.

title: “SSM Novel Applications for Agent Memory”

Research Gaps

The following questions remain unanswered by the literature:

1. No SSM work on memory contradiction detection. CausalMamba detects causal structure but not semantic contradictions between memories. This is novel territory.

2. No SSM work on belief revision. The belief revision literature (Sec 8) has not been connected to SSM architectures. An SSM that maintains a “belief state” and revises it when contradicting evidence arrives would be genuinely novel.

3. No SSM work on agent identity coherence. Multiple SSM hidden states representing the same agent across sessions — how to merge them into a single coherent identity — is unexplored.

4. No SSM work on privacy-preserving memory. SSM hidden states compress user data. What information leaks from the hidden state? Can differential privacy be applied to SSM state updates?

5. No SSM work on federated memory learning. VISION.md describes federated learning for the base model. Can SSM weights be aggregated across users without exposing individual patterns?

These gaps represent opportunities for Signet to publish novel research.