SSM-like Memory in Agent Swarms
Research incentives around world models involve throughputting huge amounts of information accurately with extremely low latency. One of the major ways world models accomplish this is by sacrificing the size of their context window - with fewer tokens to attend to, models can process attention much more efficiently, at the cost of forgetting anything they haven't seen for two seconds1. A solution to this problem has been State-Space Models (SSMs) such as Mamba2, which extend a model’s temporal memory using “state-vectors” that update on every autoregressive frame prediction3. Low latency, compute-efficient memory has not been an imminent need for LLMs, because natural methods of text compaction, .md referencing, and substantial token contexts accommodate most use cases an LLM should serve. However, this is not sufficient for agent swarms, whose individual agents act on large independent contexts and whose activity, at scale, is prone to redundancy and miscommunication4.
In our approach we are inspired by stigmergy, a mechanism of indirect communication ants use by releasing pheromones. This method allows them to organize through a shared environment, instead of direct messaging or central planning. By giving agents the parallel ability to “sense” the activity of other acting agents, our swarms can organize themselves more efficiently5.
Most approaches to modern multi-agent systems involve the coordination of agent ‘characters’. (Cognition)6 For example, one agent - “the programmer” might generate code and create a PR, while another agent - “the reviewer” engages in adversarial code review. This interaction usually goes through a dynamic workflow - a set of protocols that programmatically determine how the agents relate to each other. They may enter into a dialogue and determine what to change or whether the PR is worth merging. Because of the variance of the two agents’ contexts, and the reduced context of the reviewer, it becomes more likely that subtle issues are detected, and risks hidden in imperfect code are reduced. But every time a new generation of models is released, these workflows risk becoming brittle or obsolete. A more capable model may not need code review at all, and so the structure becomes unnecessary, or in some cases damaging to the performance of the agents. It also isn’t scalable, and restricts the ability of agents to organically order themselves in an emergent way, which is the primary benefit of stigmergic systems. Our bet is that allowing the relations between multiple agents to be trained alongside the agents themselves will allow for efficient and emergent collaboration.
The relation between agents, the essential substrate of multi-agent work, is their shared memory. In a recent position paper, Yu et al. (2026)7 argue that the most pressing open problem associated with multi-agent memory is consistency, which they break into two requirements: “update-time visibility and ordering” - when an agent’s writes become observable to other agents, and “read-time conflict resolution” - how an agent should reconcile conflicting or stale information. We aim to solve both of these problems by fine-tuning open-weight models to concurrently update global state with every token they predict, and read it whenever they make a decision. The global state maintains concurrency via an SSM-style memory latent that our models are trained to update and read from. This acts as the “trace” for agents to identify what other agents have been up to, similar to how ants leave traces as pheromones to help others organize. Most research seems to expect that multi-agent swarms will require better data hierarchies and protocols8,9. But at low latency and with a stigmergic framing, we believe a continuously updated state should be sufficient to “guide” the decisions multi-agent swarms are making together.
Early experiments are promising. We trained a 2B-parameter Qwen3-VL model to write and read a compressed memory latent10,11 that holds one token per file, using about 1.7x fewer tokens than the full shared status file. Asked multiple-choice coordination questions (which agent holds a file, which file is free, what an agent will do next), our agent answered correctly 94% of the time across three training runs, against 99.9% when reading the full status file and 38% with no shared memory. Next, we will look to prove that our shared memory can help guide agents in their decision-making, and how this approach might scale effectively.
References
- Valevski et al. (2024). Diffusion Models Are Real-Time Game Engines. arxiv.org/abs/2408.14837
- Gu and Dao (2023). Mamba: Linear-Time Sequence Modeling with Selective State Spaces. arxiv.org/abs/2312.00752
- Po et al. (2025). Long-Context State-Space Video World Models. arxiv.org/abs/2505.20171
- Cemri et al. (2025). Why Do Multi-Agent LLM Systems Fail? arxiv.org/abs/2503.13657
- Pal et al. (2026). SwarmWorld: Stigmergic technological evolution in societies of language-model agents. arxiv.org/abs/2608.26081
- Cognition (2026). Multi-Agents: What’s Actually Working. cognition.com/blog/multi-agents-working
- Yu et al. (2026). Multi-Agent Memory from a Computer Architecture Perspective: Visions and Challenges Ahead. arxiv.org/abs/2603.10062
- Han and Zhang (2025). Exploring Advanced LLM Multi-Agent Systems Based on Blackboard Architecture. arxiv.org/abs/2507.01701
- Zhang et al. (2026). PatchBoard: Schema-Grounded State Mutation for Reliable and Auditable LLM Multi-Agent Collaboration. arxiv.org/abs/2605.29313
- Ge et al. (2023). In-context Autoencoder for Context Compression in a Large Language Model. arxiv.org/abs/2307.06945
- Mu et al. (2023). Learning to Compress Prompts with Gist Tokens. arxiv.org/abs/2304.08467