title: “State-Space Models for Agent Memory Systems: Literature Review”

State-Space Models for Agent Memory Systems: Comprehensive Literature Review

This document surveys the state-space model (SSM) landscape with specific focus on applicability to Signet’s local-first agent memory system. The review covers foundational architectures, small/efficient variants, memory-specific applications, comparative analysis with transformers, training on personal data, and 2025-2026 developments.

title: “State-Space Models for Agent Memory Systems: Literature Review”

1. Foundational SSM Papers

1.1 HiPPO: Recurrent Memory with Optimal Polynomial Projections

• Authors: Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, Christopher Re
• Date: August 2020 (NeurIPS 2020)
• URL: https://arxiv.org/abs/2008.07669
• Key Insight: Introduces a general framework for online compression of continuous signals by projection onto polynomial bases. HiPPO produces operators that project arbitrary functions onto the space of polynomials, enabling optimal online memorization. The key matrices (HiPPO-LegS, HiPPO-LagT) form the mathematical backbone of all subsequent SSMs.
• Signet Relevance: HiPPO formalizes exactly what Signet’s memory system needs — a principled way to compress a continuous stream of observations (agent interactions) into a fixed-size representation that optimally preserves information about the past. The polynomial projection framework could inform how memory embeddings decay and which memories get retained.

1.2 S4: Efficiently Modeling Long Sequences with Structured State Spaces

• Authors: Albert Gu, Karan Goel, Christopher Re
• Date: November 2021 (ICLR 2022 Oral)
• URL: https://arxiv.org/abs/2111.00396
• Code: https://github.com/state-spaces/s4
• Key Insight: Proposes the Structured State Space sequence model based on a new parameterization for the SSM. S4 combines continuous-time modeling (handles irregular sampling), recurrent processing (unbounded context with constant state), and convolutional computation (efficient parallel training). It was the first model to solve the Path-X task on Long Range Arena (sequences of length 16,384), setting substantial SotA on every LRA task.
• Architecture: Continuous-time SSM discretized via bilinear method. State matrix A initialized with HiPPO. Diagonal-plus-low-rank parameterization enables O(N log N) computation via Cauchy kernel.
• Signet Relevance: S4’s ability to handle very long sequences with constant memory is directly applicable to processing long agent session histories. Its continuous-time nature means it naturally handles irregularly-spaced events (agent interactions don’t happen at fixed intervals).

1.3 H3: Hungry Hungry Hippos — Towards Language Modeling with State Space Models

• Authors: Daniel Y. Fu, Tri Dao, Khaled K. Saab, Arman W. Thomas, Atri Rudra, Christopher Re
• Date: December 2022 (ICLR 2023)
• URL: https://arxiv.org/abs/2212.14052
• Code: https://github.com/HazyResearch/H3
• Key Insight: Identified two critical capabilities SSMs lacked for language modeling: (1) recalling earlier tokens in the sequence, and (2) comparing tokens across the sequence. H3 addresses these by combining an SSM layer with a multiplicative interaction (similar to attention’s QKV pattern). This was the first SSM to approach transformer quality on language modeling and scale to billions of parameters.
• Architecture: Two SSM layers (shift SSM for copying, diagonal SSM for comparison) with multiplicative gating. Critical component: hardware efficiency through FlashConv, which is essential for scaling SSMs to billion-parameter models.
• Signet Relevance: H3’s insight about recall vs. comparison maps directly to memory retrieval needs. An agent memory system needs both capabilities: recalling specific past interactions (the shift/copy mechanism) and comparing current context against stored memories (the comparison mechanism).

1.4 Hyena Hierarchy: Towards Larger Convolutional Language Models

• Authors: Michael Poli, Stefano Massaroli, Eric Nguyen, Daniel Y. Fu, Tri Dao, Stephen Baccus, Yoshua Bengio, Stefano Ermon, Christopher Re
• Date: February 2023 (ICML 2023)
• URL: https://arxiv.org/abs/2302.10866
• Key Insight: Proposes a subquadratic drop-in replacement for attention constructed by interleaving implicitly parametrized long convolutions and data-controlled gating. Hyena achieves attention-quality results with subquadratic cost. At sequences of hundreds of thousands of tokens, Hyena operators are over 100x faster than optimized attention.
• Architecture: Order-N Hyena operator: N rounds of element-wise gating interleaved with long convolutions. Implicit parameterization of filters via a small neural network. Data-controlled gating provides the selectivity that purely linear models lack.
• Signet Relevance: Hyena’s implicit filter parameterization is elegant for memory systems — the filters themselves are learned and can adapt, rather than being fixed. The data-controlled gating is similar to what Signet needs: selectively attending to relevant memories based on current context.

1.5 Mamba: Linear-Time Sequence Modeling with Selective State Spaces

• Authors: Albert Gu, Tri Dao
• Date: December 2023
• URL: https://arxiv.org/abs/2312.00752
• Code: https://github.com/state-spaces/mamba
• Key Insight: The breakthrough paper. Mamba introduces selective SSMs where the (A, B, C, delta) parameters are functions of the input, allowing the model to selectively propagate or forget information along the sequence depending on the current token. Combined with a hardware-aware parallel algorithm and a simplified architecture (no attention, no MLP blocks), Mamba achieves 5x throughput vs transformers, linear scaling in sequence length, and Mamba-3B matches transformers twice its size.
• Three key ingredients (per Albert Gu’s July 2025 analysis):
  1. State size: Hidden state h_t is N-dimensional (N >> 1), storing N times more information than classical RNNs like LSTMs.
  2. State expressivity: Selective (data-dependent) state transitions A_t, B_t, C_t allow the model to choose what to remember vs forget — impossible with time-invariant SSMs.
  3. Training efficiency: Hardware-aware parallel scan algorithm enables efficient GPU training despite the recurrent formulation.
• Signet Relevance: CRITICAL. Mamba’s selective mechanism is essentially what Signet’s memory scorer needs — a learned function that decides what information to retain and what to forget, operating in constant memory with linear-time complexity. The 5x throughput advantage matters for a sidecar process that must not impact the host system.

1.6 Mamba-2: Transformers are SSMs (State Space Duality)

• Authors: Albert Gu, Tri Dao
• Date: May 2024
• URL: https://arxiv.org/abs/2405.21060
• Key Insight: Proves a deep mathematical connection: the SSD (Structured State Space Dual) model is simultaneously an SSM and a form of linear attention. The duality shows that a scalar-identity structure on A_t makes the SSM equivalent to causal linear attention with an input-dependent positional mask. This enables much larger state dimensions (N=64 to N=256 vs N=16 in Mamba-1) while being much faster to train via matrix multiplications.
• Key tradeoff: Mamba-2 restricts the diagonal A to scalar-times- identity (less expressive per-dimension dynamics) but compensates with much larger state size and faster training. Pure inference may still favor Mamba-1’s more expressive diagonal A at smaller state sizes.
• Performance: At the Pile benchmark, slightly better scaling laws than Mamba-1. Substantially better on hard associative recall tasks (MQAR) due to 16x larger state sizes.
• Signet Relevance: The SSD algorithm’s “chunkwise” processing (splitting sequences into segments, computing attention-like operations within segments, passing state between segments) maps well to how agent sessions work — you process a session chunk, carry state forward.

1.7 Mamba-3: Improved Sequence Modeling using State Space Principles

• Authors: Aakash Lahoti, Kevin Y. Li, Berlin Chen, Caitlin Wang, Aviv Bick, J. Zico Kolter, Tri Dao, Albert Gu
• Date: March 16, 2026 (3 days ago)
• URL: https://arxiv.org/abs/2603.15569
• Affiliations: CMU, Princeton, Together AI, Cartesia AI
• Key Insight: Three methodological improvements driven by an “inference-first” perspective:
  1. Exponential-Trapezoidal Discretization: More expressive recurrence derived from proper SSM discretization theory. Reveals an implicit convolution in the SSM input, replacing the short causal convolution previously thought essential for recurrent models.
  2. Complex-valued State: Enables richer state tracking (solves parity/arithmetic tasks that Mamba-2 cannot). Equivalent to a data-dependent rotary embedding, efficiently computable.
  3. Multi-Input Multi-Output (MIMO): Switches from outer-product to matrix-multiplication state updates. Increases decoding FLOPs by 4x at fixed state size while maintaining similar wall-clock latency, improving both performance and hardware utilization.
• Results at 1.5B: +2.2 over Transformers, +1.9 over Mamba-2, +1.8 over Gated DeltaNet on downstream accuracy. Achieves Mamba-2 perplexity with half the state size.
• Signet Relevance: The MIMO formulation is directly relevant — it lets you pack more computation into the memory-bound decode step without increasing state size or latency. For a sidecar process doing inference, this means better predictions at the same memory cost. The complex- valued state tracking is interesting for modeling cyclical patterns (daily/weekly usage patterns).

title: “State-Space Models for Agent Memory Systems: Literature Review”

2. The Broader Modern Recurrent Model Family

2.1 Related Architectures

All of the following models can be cast into the same SSM equation h_t = A_t * h_{t-1} + B_t * x_t, differing mainly in the structure of A_t and the training algorithm:

Critical observation from Albert Gu (July 2025): “All of these models are much more similar to each other than they are to quadratic attention.” The real divide is between constant-memory compressed-state models (SSMs/linear RNNs) and growing-cache models (transformers with KV cache).

2.2 Test-Time Training (TTT) Layers

• Paper: “Learning to (Learn at Test Time): RNNs with Expressive Hidden States”
• Authors: Yu Sun, Xinhao Li, Karan Dalal, et al.
• Date: July 2024
• URL: https://arxiv.org/abs/2407.04620
• Key Insight: Makes the hidden state of a recurrent model a machine learning model itself, and the update rule a step of self-supervised learning. TTT-Linear (hidden state = linear model) and TTT-MLP (hidden state = 2-layer MLP). Unlike Mamba which plateaus after 16k context, TTT keeps reducing perplexity with more tokens, like transformers.
• Signet Relevance: HIGH. TTT directly formalizes what Signet’s predictor does conceptually — learning from the test sequence itself. The hidden state as an “associative memory updated by online optimization” is exactly the paradigm of learning user patterns at inference time.

2.3 Hybrid SSM-Transformer Models

The consensus from 2024-2026 research is that the optimal architecture combines SSM layers with a small proportion of attention layers:

Optimal ratio: 3:1 to 10:1 SSM:attention layers, independently verified by dozens of research groups. This is not just about efficiency — it actually improves modeling power. “Taking a pure Transformer and replacing most layers with SSM layers would both improve efficiency and performance.” (Gu, July 2025)

title: “State-Space Models for Agent Memory Systems: Literature Review”

3. Small/Efficient SSM Architectures

3.1 IBM FlowState (9.1M Parameters)

• Organization: IBM Research
• Date: September 2025 (NeurIPS 2025 Workshop)
• URL: https://research.ibm.com/blog/SSM-time-series-model
• Model: https://huggingface.co/ibm-granite/granite-timeseries-flowstate-r1
• Key Insight: At just 9.1M parameters, FlowState is the smallest model in the GIFT-Eval top 10 for zero-shot time-series forecasting, outperforming rivals over 20x its size. Uses an S5-based encoder paired with a basis-function decoder that converts timescale-invariant hidden states into predictions at arbitrary resolution.
• Architecture: S5 SSM encoder + basis-function decoder. The encoder converts time series into timescale-invariant representations; the decoder interprets hidden state elements as coefficients of basis functions.
• Signet Relevance: CRITICAL. This proves that SSMs at sub-10M parameters can outperform much larger models on temporal prediction tasks. This is exactly the parameter range for a Rust sidecar binary. The timescale-invariant encoding is relevant for modeling agent interactions that happen at varying frequencies.

3.2 Quantizing Small-Scale SSMs for Edge AI

• Authors: Leo Zhao, Tristan Torchet, Melika Payvand, Laura Kriener, Filippo Moro
• Date: June 2025
• URL: https://arxiv.org/abs/2506.12480
• Key Insight: Analyzes quantization effects on small-scale SSMs (S4D architecture). Post-training quantization (PTQ) at 8-bit drops performance to 40% on sMNIST, but quantization-aware training (QAT) recovers it to 96%. State matrix A and internal state x are particularly sensitive to quantization. Proposes heterogeneous quantization strategy (different precision for different components) achieving 6x memory reduction without performance loss.
• Key finding: The A matrix and hidden state require higher precision than weights, suggesting mixed-precision is essential for edge SSMs.
• Signet Relevance: DIRECTLY APPLICABLE. For the Rust sidecar, this paper provides the recipe: use QAT (not PTQ), keep A matrix and hidden state at higher precision (fp16 or bf16), quantize weights to int8 or lower. 6x memory reduction means a 10M-parameter SSM could fit in ~7MB of memory.

3.3 Efficient Unstructured Pruning of Mamba for Edge

• Authors: Ibne Farabi Shihab, Sanjeda Akter, Anuj Sharma
• Date: November 2025 (EMNLP 2025)
• URL: https://aclanthology.org/2025.emnlp-main.562/
• Key Insight: Achieves up to 70% parameter reduction on Mamba with only 3-9% performance drop. Uses pruning based on both weight and gradient importance (preserving Mamba’s unique recurrent dynamics), gradual pruning schedule, and global allocation strategy. Results: 1.77x faster inference, 46% memory reduction.
• Key finding: Mamba is remarkably robust to pruning, more so than transformers. This is likely because the recurrent state naturally distributes information across parameters rather than concentrating it in specific attention heads.
• Signet Relevance: DIRECTLY APPLICABLE. A 130M Mamba model pruned 70% becomes ~39M parameters. Combined with int8 quantization, this could yield a sub-20MB binary — well within sidecar constraints.

3.4 Distilling Low-Rank Mamba for Edge

• Paper: “Distilling Low-Rank Mamba for Edge Multispectral Fusion Object Detection”
• Date: March 2026
• URL: https://arxiv.org/abs/2603.06920
• Key Insight: Applies knowledge distillation and low-rank factorization to create lightweight Mamba variants for edge deployment. Demonstrates that Mamba’s state structure is amenable to rank reduction without catastrophic quality loss.

3.5 Parameter Budget Analysis for Signet

Given the research, here is a realistic parameter budget for a Signet memory predictor SSM sidecar:

The sweet spot for a Rust sidecar appears to be 2-10M parameters, yielding a binary in the 5-20MB range. IBM’s FlowState proves this range is viable for strong temporal prediction.

title: “State-Space Models for Agent Memory Systems: Literature Review”

4. SSMs for Memory and Retrieval

4.1 Mathematical Formalism for Memory Compression in Selective SSMs

• Author: Siddhanth Bhat
• Date: October 2024
• URL: https://arxiv.org/abs/2410.03158
• Key Insight: Develops rigorous information-theoretic framework for understanding memory compression in selective SSMs. Uses mutual information and rate-distortion theory to formalize the tradeoff between memory efficiency and information retention. Proves stability and convergence of hidden state in selective SSMs, ensuring reliable long-term memory retention. Provides theoretical bounds on compressible information without sacrificing performance.
• Key results: Selective gating mechanism (Mamba’s selectivity) dynamically filters based on input relevance, achieving significant compression. The rate-distortion analysis shows SSMs operate near the theoretical optimum for sequence compression.
• Signet Relevance: HIGH. This paper provides the theoretical foundation for using SSMs as memory compressors. The information- theoretic bounds tell us exactly how much agent interaction history can be compressed into a fixed-size state without losing critical information for prediction.

4.2 SleepGate: Sleep-Inspired Memory Consolidation for LLMs

• Author: Ying Xie (Kennesaw State University)
• Date: March 15, 2026 (5 days ago)
• URL: https://arxiv.org/abs/2603.14517
• Key Insight: Proposes biologically-inspired framework for active memory management in transformer KV caches, directly inspired by sleep-dependent memory consolidation. Three mechanisms:
  1. Conflict-aware temporal tagger: Detects when new entries supersede old ones using semantic signatures.
  2. Forgetting gate: Lightweight network trained to selectively evict or compress stale cache entries.
  3. Consolidation module: Merges related surviving entries into compact summary representations. These operate in “sleep micro-cycles” triggered by attention entropy.
• Results: 99.5% retrieval accuracy at interference depth 5, 97% at depth 10, while all baselines remain below 18%.
• Connection to SSMs (from paper): “State-space models implicitly implement forms of forgetting through their inductive biases. However, these are fixed architectural choices rather than learned, content- dependent forgetting policies.”
• Signet Relevance: DIRECTLY RELEVANT TO SIGNET’S RETENTION DECAY. SleepGate’s three modules map exactly to Signet’s memory pipeline needs: (1) detecting when new memories supersede old ones (entity dedup), (2) selectively forgetting stale memories (retention decay), (3) consolidating related memories into summaries (session summaries). The “sleep micro-cycle” concept could be a periodic maintenance pass.

4.3 The Brain-Database Analogy (Gu, July 2025)

Albert Gu’s influential blog post “On the Tradeoffs of SSMs and Transformers” (https://goombalab.github.io/blog/2025/tradeoffs/) provides the most useful conceptual framework for understanding SSMs in a memory context:

• Transformers are like databases: Every observation is filed away for future reference (KV cache grows linearly). Good for exact retrieval of specific items.
• SSMs are like brains: Finite-sized memories that are always on, processing new inputs in real-time. Good for maintaining a long, persistent summary of context without needing to recall every detail.

“Maintaining a continual conversation with an assistant is much more like human conversations and relationships: what matters is a long, persistent summary of the context, remembering the shape and flow of the interactions without needing to recall every specific detail. No one needs a scratchpad to have a continual relationship with their friend. This is exactly where the more brain-like nature of SSMs is more suitable.”

This is exactly Signet’s use case. The agent memory system needs the “brain-like” persistent compressed state for ongoing context, augmented by database-like retrieval (the SQLite + FTS5 + vector search that already exists) for specific recall.

4.4 Infini-attention: Compressive Memory for Transformers

• Authors: Google (Tsendsuren Munkhdalai et al.)
• Date: April 2024
• URL: https://arxiv.org/abs/2404.07143
• Key Insight: Incorporates a compressive memory into vanilla attention, theoretically enabling infinite context length with bounded memory. Maintains a single buffer for all memory of earlier segments.
• Practical issues: Follow-up analysis by Hugging Face showed the approach struggles in practice, with the compressive memory failing to adequately represent earlier context.
• Signet Relevance: CAUTIONARY. Shows that naive compression of attention-based memory doesn’t work well. SSMs’ inherent compression (via the recurrent state) may be more principled than bolting compression onto attention.

4.5 Memory Augmented State Space Models for Time Series

• Source: IJCAI 2022
• URL: https://www.ijcai.org/proceedings/2022/0479.pdf
• Key Insight: Unlike fixed-order Markovian SSMs, this model features an external memory system that stores informative latent state experiences. The external memory augments the SSM’s fixed-size state with retrievable historical states.
• Signet Relevance: The hybrid of SSM compressed state + external memory retrieval is exactly the architecture Signet should consider: SSM for temporal pattern compression, SQLite/vector store for retrievable specific memories.

4.6 SSMs for Sequential Recommendation

Multiple recent papers apply SSMs to sequential recommendation, which is closely analogous to predicting memory relevance:

• SIGMA (AAAI 2025): “Selective Gated Mamba for Sequential Recommendation” — Uses Mamba’s selective mechanism to model user preference evolution over item interaction sequences.
• M2Rec (May 2025): “Multi-scale Mamba for Efficient Sequential Recommendation” — Multi-scale SSM captures both short-term and long-term user preference patterns.
• Hierarchical Mamba for Recommendation (2025): Models long and short-term preference with hierarchical Mamba architecture.

These are directly analogous to Signet’s task: given a sequence of agent interactions, predict which stored memories will be relevant to the next interaction.

title: “State-Space Models for Agent Memory Systems: Literature Review”

5. SSMs vs Transformers: Where SSMs Clearly Win

5.1 Summary from Comprehensive Benchmarking (July 2025)

From “Characterizing State Space Model (SSM) and SSM-Transformer” (https://arxiv.org/abs/2507.12442):

5.2 Key Quantitative Findings

From Albert Gu’s analysis (July 2025) and NVIDIA’s benchmarks:

1. Throughput: Mamba achieves 5x higher throughput than transformers of equal size at inference time.
2. Memory scaling: Transformer KV cache grows linearly with context; SSM state is constant regardless of context length.
3. Byte-level modeling: When comparing on raw bytes (no tokenization), SSMs strongly outperform transformers even when transformers use 2x more FLOPs. “Keeping the same models and same data, but simply untokenizing the inputs, simultaneously lets the Transformer use much more compute but also decreases its performance relative to the SSM.”
4. DNA modeling: SSMs scale substantially better than transformers on DNA (4-token vocabulary, high-resolution input).
5. Hybrid advantage: Replacing 70-90% of transformer layers with SSM layers improves both efficiency AND quality.

5.3 Where SSMs Struggle

1. Exact associative recall: “SSMs can’t memorize a phonebook in one pass and then recite it back.” But Mamba-3’s complex-valued state and MIMO formulation significantly improve recall capabilities.
2. State tracking: Pure SSMs historically failed at parity detection of bit sequences. Mamba-3’s complex-valued state solves this.
3. In-context learning: SSMs show weaker in-context learning than transformers in some settings, though TTT layers address this.

5.4 The Tokenization Insight

Gu’s analysis reveals a deeper principle: “The inductive bias of soft attention is hard attention.” Transformers work best on pre-compressed, semantically meaningful tokens. When data is high-resolution and individual elements aren’t meaningful (characters, raw sensor data, time-series points), SSMs have a clear modeling advantage because they naturally compress data into meaningful abstractions.

For Signet: agent interaction data (timestamps, memory accesses, feedback signals) is high-resolution, irregularly-sampled, and individual data points are not independently meaningful. This is exactly the regime where SSMs excel.

title: “State-Space Models for Agent Memory Systems: Literature Review”

6. Training SSMs on Small/Personal Data

6.1 Few-Shot and Low-Data Properties

SSMs have inherent advantages for low-data regimes:

1. Structural priors: The SSM parameterization (A matrix from HiPPO, continuous-time formulation) provides strong inductive biases that reduce sample complexity. As noted in the Deep State Space Models for Time Series paper (NeurIPS): “When there is little data to learn from, the structure imposed by the SSM can alleviate overfitting.”

2. FlowState proof point: IBM’s 9.1M-parameter FlowState achieves top-10 zero-shot performance on GIFT-Eval, outperforming 200M+ parameter models. This demonstrates that SSMs can achieve strong generalization with far fewer parameters, implying less training data is needed.

3. Transfer from pre-training: The HiPPO initialization provides a mathematically optimal starting point for the state dynamics. This is a form of inductive bias that doesn’t require pre-training data.

6.2 Online/Continual Learning

Several works directly address continual learning with SSMs:

1. TTT layers (Sun et al., 2024): The hidden state literally learns from the test sequence via self-supervised gradient steps. This is the most direct form of online learning — the model adapts to each new user/context at inference time.

2. TTT4Rec (September 2024, https://arxiv.org/abs/2409.19142): Applies test-time training specifically to recommendation, using self-supervised learning during inference to dynamically update model parameters. Directly applicable to Signet’s memory relevance prediction.

3. FedSSM (IEEE TMC, 2026): Uses state-space models to mitigate catastrophic forgetting in personalized federated learning. By capturing temporal evolution of model parameters through hidden states, SSMs enhance retention of critical knowledge across training rounds. This directly validates SSMs for personalization without forgetting.

4. STAD (October 2024, https://arxiv.org/abs/2407.12492): Temporal Test-Time Adaptation with State-Space Models. Proposes a probabilistic SSM that adapts deployed models to temporal distribution shifts by learning the dynamics of distribution change.

6.3 Personalization Strategy for Signet

Based on the research, the optimal strategy for Signet would be:

1. Pre-train a base SSM on anonymized community training signals (the federated learning approach described in VISION.md).
2. Ship the base model with every Signet install (9-25M parameters, <20MB binary).
3. Fine-tune locally using the user’s own interaction data via:
   • Low-rank adaptation (LoRA-like) to keep fine-tuning efficient
   • Online learning with the user’s own memory feedback signals
   • Periodic “sleep” consolidation passes (SleepGate-inspired)
4. The SSM hidden state itself serves as a persistent representation of the user’s interaction patterns, updated with each session.

This maps 1:1 to VISION.md: “A neural network unique to each user, trained on their own interaction patterns, that gets sharper the longer you use it. No shared personal weights. Your weights never leave your machine.”

title: “State-Space Models for Agent Memory Systems: Literature Review”

7. Recent Developments (2025-2026)

7.1 Production Deployments

SSMs have moved from research to production:

• Cartesia AI (Albert Gu’s company): Production SSM-based models for real-time audio (Sonic TTS) and language. Claims models efficient enough to “run pretty much anywhere.” Uses SSMs for streaming voice synthesis in 40+ languages.
• IBM Granite 4.0: Incorporating Bamba’s hybrid SSM-transformer architecture into production enterprise models.
• NVIDIA Nemotron-H: Production hybrid at 560B total parameters with state-of-the-art performance.
• Qwen3 / Kimi Linear: Production LLMs using Gated DeltaNet (SSM family) layers.
• Together AI: Hosting Mamba-3 models for inference, actively developing SSM infrastructure.

7.2 Rust/Native Inference Ecosystem

The Rust ML ecosystem now supports SSM inference:

• Candle (Hugging Face): Minimalist ML framework for Rust with Mamba inference support. “Serverless (on CPU), small and fast deployments.” Supports quantization. GitHub: https://github.com/huggingface/candle
• Burn: Rust ML framework with ONNX import support.
• ONNX Runtime: Mature Rust bindings, supports custom ops needed for SSM selective scan.

Candle’s existing Mamba implementation is particularly relevant — it demonstrates that SSM inference in Rust on CPU is already viable today. This could serve as the foundation for Signet’s predictor sidecar.

7.3 Timeline of Key Developments

title: “State-Space Models for Agent Memory Systems: Literature Review”

8. Synthesis: Architecture Recommendation for Signet

8.1 Why SSMs Are the Right Choice for Signet’s Predictor

1. Constant memory: The predictor processes arbitrarily long interaction histories without growing memory usage.
2. Linear time: Processing cost scales linearly with session length, not quadratically.
3. Streaming inference: SSMs naturally process data as a stream, matching how agent interactions arrive in real-time.
4. Compression is the point: The SSM hidden state IS a compressed representation of interaction history. This is not a limitation — it is the desired behavior for a memory relevance predictor.
5. Small footprint: Proven at 9.1M parameters (FlowState). Amenable to quantization (6x reduction) and pruning (70% sparsity).
6. Rust-ready: Candle already implements Mamba inference in Rust.
7. Online learning: TTT and related work show SSMs can adapt at inference time, enabling the personalization described in VISION.md.

8.2 Recommended Architecture

Signet Memory Predictor SSM
=============================
Architecture: Mamba-2 or Mamba-3 based (selective SSM)
Parameters:   5-10M (4-8 layers, d_model=192-256, N=32-64)
Precision:    Heterogeneous (A matrix fp16, weights int8)
Binary size:  ~10-15MB (Rust, quantized)
Inference:    CPU-only, <5ms per interaction event
Training:     Online fine-tuning via LoRA + periodic consolidation
State:        Persistent hidden state saved between sessions

Input features per interaction event:
- Memory access patterns (which memories were retrieved)
- Agent feedback signals (helpful/not helpful)
- Temporal features (time-of-day, day-of-week, session duration)
- Content similarity scores (current context vs. retrieved memories)
- Memory metadata (age, access count, decay score)

Output:
- Memory relevance scores (which memories to inject)
- Retention predictions (which memories to preserve vs. decay)
- Temporal pattern indicators (cyclical access patterns)

8.3 Open Questions for Further Research

1. Complex-valued state for cyclical patterns: Mamba-3’s complex state could naturally model daily/weekly usage patterns. Worth prototyping.
2. MIMO for multi-signal prediction: Can MIMO SSMs simultaneously predict relevance, retention, and temporal patterns from a single forward pass?
3. SleepGate for memory consolidation: Can the sleep micro-cycle concept replace or augment Signet’s current retention decay?
4. Federated pre-training: What’s the minimum community data needed to pre-train a useful base model? FlowState’s success at 9.1M params suggests the data requirements may be modest.
5. SSM state as identity embedding: Could the persistent SSM state serve as a compact representation of agent identity/personality, complementing SOUL.md?

title: “State-Space Models for Agent Memory Systems: Literature Review”

References

Complete bibliography of papers cited in this review:

1. Gu, A., Dao, T., Ermon, S., Rudra, A., Re, C. (2020). HiPPO: Recurrent Memory with Optimal Polynomial Projections. NeurIPS 2020. arXiv:2008.07669
2. Gu, A., Goel, K., Re, C. (2022). Efficiently Modeling Long Sequences with Structured State Spaces. ICLR 2022. arXiv:2111.00396
3. Fu, D.Y., Dao, T., Saab, K.K., Thomas, A.W., Rudra, A., Re, C. (2023). Hungry Hungry Hippos: Towards Language Modeling with State Space Models. ICLR 2023. arXiv:2212.14052
4. Poli, M. et al. (2023). Hyena Hierarchy: Towards Larger Convolutional Language Models. ICML 2023. arXiv:2302.10866
5. Gu, A., Dao, T. (2023). Mamba: Linear-Time Sequence Modeling with Selective State Spaces. arXiv:2312.00752
6. Gu, A., Dao, T. (2024). Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality. arXiv:2405.21060
7. Lahoti, A. et al. (2026). Mamba-3: Improved Sequence Modeling using State Space Principles. arXiv:2603.15569
8. Smith, J.T., Warrington, A., Linderman, S.W. (2023). Simplified State Space Layers for Sequence Modeling. ICLR 2023. arXiv:2208.04933
9. Sun, Y. et al. (2024). Learning to (Learn at Test Time): RNNs with Expressive Hidden States. arXiv:2407.04620
10. Bhat, S. (2024). Mathematical Formalism for Memory Compression in Selective State Space Models. arXiv:2410.03158
11. Xie, Y. (2026). Learning to Forget: Sleep-Inspired Memory Consolidation for Resolving Proactive Interference in Large Language Models. arXiv:2603.14517
12. Zhao, L. et al. (2025). Quantizing Small-Scale State-Space Models for Edge AI. arXiv:2506.12480
13. Shihab, I.F., Akter, S., Sharma, A. (2025). Efficient Unstructured Pruning of Mamba State-Space Models for Resource-Constrained Environments. EMNLP 2025.
14. AI21 Labs (2024). Jamba: A Hybrid Transformer-Mamba Language Model. arXiv:2403.19887
15. NVIDIA (2025). Nemotron-H: A Family of Accurate, Efficient Hybrid Mamba-Transformer Models. arXiv:2504.03624
16. IBM Research (2025). Bamba: An Open Hybrid SSM-Transformer Model. Hugging Face.
17. IBM Research (2025). FlowState: SSM-Based Time-Series Foundation Model. NeurIPS 2025 Workshop.
18. Zhang, W. et al. (2026). Mitigating Catastrophic Forgetting in Personalized Federated Learning for Edge Devices Using State-Space Models. IEEE TMC.
19. Gu, A. (2025). On the Tradeoffs of SSMs and Transformers. Goomba Lab Blog.
20. Munkhdalai, T. et al. (2024). Efficient Infinite Context Transformers with Infini-attention. arXiv:2404.07143
21. Beck, M. et al. (2024). xLSTM: Extended Long Short-Term Memory. NeurIPS 2024. arXiv:2405.04517
22. Yang, S. et al. (2024). Gated Linear Attention / DeltaNet. arXiv:2406.06484
23. SIGMA (2025). Selective Gated Mamba for Sequential Recommendation. AAAI 2025.
24. M2Rec (2025). Multi-scale Mamba for Efficient Sequential Recommendation. arXiv:2505.04445
25. Cartesia AI (2024). Based: Simple Linear Attention Language Models Balance the Recall-Throughput Tradeoff.
26. Graf, L. et al. (2025). FlowState-R1: A State-Space Model for Time-Series Forecasting. IBM Research.