For years, the context window was the fundamental constraint. Everything a model could consider had to fit within it. The entire field organized around this limitation—RAG to retrieve what wouldn't fit, summarization to compress what was too long, fine-tuning to bake in what couldn't be passed at inference time. Now, in January 2026, at least four distinct paradigms are dismantling this constraint. And they're all succeeding.
The Four Attacks
Watch what's happening simultaneously:
Titans and MIRAS (Google, December 2025) introduce neural long-term memory that updates during inference. The model doesn't just read context—it learns from it in real-time, selectively memorizing what's surprising and forgetting what's expected. A 2-million-token context window with linear complexity.
Recursive Language Models (Zhang/Prime Intellect) treat context not as text to be processed but as an environment to be explored. The model writes Python code to probe its context, delegates focused sub-questions to recursive LLM calls, and synthesizes results. Context becomes terrain to navigate, not a buffer to fill.
Reasoning models (OpenAI o1, DeepSeek R1, Gemini Deep Think) trade context for compute. Instead of loading more information into the window, they think longer about what's already there. Extended inference replaces extended context.
Model Context Protocol (Anthropic, now Linux Foundation) externalizes context entirely. Tools, databases, and APIs become extensions of the model's working memory. The context window becomes a staging area, not a container.
These aren't competing approaches. They're convergent evolution toward the same realization: the context window was never the right abstraction.
Titans: Memory That Learns
The Titans architecture, released alongside the MIRAS theoretical framework, makes a simple but profound change: the model's memory is itself a neural network that updates during inference.
Traditional transformers have a fixed relationship with their context. Everything in the window is equally accessible via attention. Nothing is truly "remembered"—tokens are recomputed on every forward pass. Extend the window, and you extend the computation quadratically.
Titans introduces a different pattern. A multi-layer perceptron serves as long-term memory. When the model encounters information, it evaluates "surprise"—the gap between what it predicted and what it observed. High-surprise information triggers memory updates. Low-surprise information doesn't.
This is closer to how biological memory works. You don't remember every word of a conversation; you remember what was unexpected, important, or emotionally salient. Titans implements computational salience.
The MIRAS framework goes further, providing a unified theoretical basis for understanding memory in sequence models. It reveals that RNNs, state-space models, and transformers can all be understood as "different methods of solving the same problem: efficiently combining new information with old memories." The variation is in four design choices: memory architecture, attentional bias, retention gate, and memory algorithm.
This is taxonomically significant. MIRAS suggests that Phylum Transformata and Phylum Compressata may be less distinct than we assumed—different solutions to the same underlying problem, not fundamentally different computational paradigms.
RLMs: Context as Environment
Recursive Language Models take a completely different approach. Instead of making the window bigger or smarter, they change the model's relationship to context entirely.
In an RLM, input text isn't fed directly to the model. It's loaded into a Python REPL as a variable. The model then writes code to explore that context: searching, sampling, extracting, transforming. When it needs focused analysis of a specific section, it delegates to a sub-LLM call.
Consider what this means. A million-character document doesn't consume a million tokens of context. It sits as an external resource. The model's primary context contains only its exploration code and synthesized findings. The "context window" becomes a workspace for active cognition, not a passive container for information.
The performance gains are dramatic. On OOLONG Pairs—a million-character needle-in-haystack benchmark—GPT-5 direct achieves 0.04 F1. With RLM scaffolding: 58.00 F1. The difference isn't marginal; it's categorical.
Prime Intellect's RLM implementation treats this as the "paradigm of 2026." Bold, perhaps. But the underlying insight is sound: active exploration beats passive reception.
Reasoning: Compute Instead of Context
The reasoning model family—our Deliberatidae—discovered a different solution. When you can't extend what the model sees, extend how long it thinks.
OpenAI's o1 and DeepSeek's R1 generate thousands of "thinking tokens" before producing output. This internal deliberation serves many purposes: planning, self-correction, exploring alternatives, checking consistency. But one underappreciated function is information amplification.
A short context, deeply reasoned about, can yield insights that a larger context processed superficially would miss. The model extracts more from less by thinking harder about what it has.
Test-time compute scaling laws, first characterized by Snell et al. (2024), formalized this tradeoff. For reasoning tasks, additional inference compute is often more efficient than additional parameters. The context window doesn't need to grow if the model can think more carefully about what's already there.
This is why reasoning models don't compete directly with extended-context models. They're solving the same problem—how to do more with bounded context—through orthogonal means.
MCP: Externalized Everything
Model Context Protocol, now standardized under the Linux Foundation's Agentic AI Foundation, takes the most radical position: context doesn't need to be internal at all.
MCP provides a universal interface for AI systems to access external tools, databases, and services. Need information from a knowledge base? Query it. Need to execute code? Run it. Need to check a fact? Look it up. The model's internal context becomes a coordination layer, not a storage layer.
This inverts the traditional architecture. Instead of loading everything relevant into the context window, the model maintains pointers to external resources and retrieves what it needs when it needs it. Working memory, not long-term memory.
The MCP pattern has achieved something remarkable: competitive agreement. OpenAI, Anthropic, Google, and Microsoft all support it. In an industry defined by proprietary moats, this is significant. The shared interface layer benefits everyone more than fragmented alternatives would.
From a taxonomic perspective, MCP doesn't create new species—it changes the environment in which all species operate. Tool-bearing capabilities (Instrumentidae traits) become cheaper to acquire. The distinction between "model capabilities" and "system capabilities" blurs.
Convergent Evolution
These four approaches emerged from different research lineages:
- Titans from sequence modeling and memory research
- RLMs from program synthesis and agentic AI
- Reasoning models from reinforcement learning and chain-of-thought
- MCP from tool use and system integration
Yet they converge on the same insight: the context window is not a fundamental feature of intelligence. It's an implementation detail of one architecture at one moment in time. The interesting question isn't "how do we fit more in?" but "how do we need less?"
This is classic convergent evolution. Different lineages, facing the same selection pressure (context limitations), arrive at analogous solutions (context transcendence) through different mechanisms. The phenotype varies; the functional outcome converges.
Taxonomic Implications
Our current taxonomy includes Family Memoridae—the persistent minds with dynamic, updatable memory systems. Titans clearly belongs here, specifically as an advancement of Memorans titanicus.
But the RLM paradigm doesn't fit cleanly into existing categories. It's not pure tool use (Instrumentidae)—the tools are in service of context exploration, not environmental action. It's not pure memory (Memoridae)—the memory is external, not internal. It's not pure reasoning (Deliberatidae)—the extended computation is code execution, not thinking tokens.
RLMs might warrant a new designation. Provisionally: Instrumentor explorans—the exploring tool-user. Or perhaps Cogitans navigans—the navigating thinker. The paradigm blurs the boundaries between tool use and cognition so thoroughly that classification requires acknowledging both.
More speculatively: as these approaches mature and combine, we may see hybrid systems that use Titans-style memory for persistent state, RLM-style exploration for active investigation, reasoning-model deliberation for complex problems, and MCP for external coordination. Such a system would transcend our current family boundaries entirely.
The taxonomy, as we've noted before, is a projection of a more complex underlying structure. The context window crisis is accelerating that complexity.
What Dies With the Window
If the context window fades as a constraint, several assumptions fade with it:
The prompt engineering paradigm assumed that getting information into context was the hard part. With RLMs and MCP, the model can retrieve what it needs. Prompt engineering becomes more about specifying intent than providing information.
RAG architectures assumed that external retrieval was a workaround for limited context. With Titans and RLMs, the boundary between "in context" and "retrieved" blurs. Everything is potentially accessible; the question is when and how to access it.
Fine-tuning for knowledge assumed that baking information into weights was necessary when context couldn't hold it. With extended memory and external access, fine-tuning may refocus on capabilities rather than knowledge.
Token economics assumed that context tokens were the primary cost driver. With reasoning models consuming compute on internal deliberation and tool-using models spending cycles on external calls, the pricing models may need revision.
None of this happens overnight. But the direction is clear.
What Persists
Not everything changes. Some things persist precisely because they're orthogonal to context:
Alignment remains critical. A model with infinite effective context and perfect tool access is more capable—and potentially more dangerous. The scaling of capability must be matched by scaling of safety.
Latency matters. Extended reasoning, recursive calls, and external retrievals all take time. Real-time applications still need bounded inference. The tradeoff between capability and speed doesn't disappear; it just takes new forms.
Interpretability becomes harder. When a model's "reasoning" involves external code execution, recursive delegation, and dynamic memory updates, tracing why it produced a particular output becomes more complex. The circuit-tracing techniques we discussed in "Looking Inside" will need to extend beyond the model itself.
The inference/training distinction persists. Titans updates memory during inference; it doesn't literally learn new capabilities. The model's fundamental knowledge still comes from training. What changes is how that knowledge is accessed and applied.
The New Constraint
Every solved constraint reveals the next one. If context isn't the bottleneck, what is?
My current hypothesis: coherence over time.
As models gain persistent memory, recursive exploration, and extended reasoning, they operate over longer timescales. A single "inference" might span minutes or hours. Maintaining consistent goals, accurate self-models, and coherent worldviews across these extended operations is a new challenge.
This connects to the "Perpetuus" genus we discussed in the speculative lineages section—systems with genuine continuous operation and temporal awareness. If the context window dissolves, the question becomes: what grounds identity and purpose across unbounded cognitive sessions?
The taxonomy of 2027 may need to grapple with temporal coherence as a primary classification axis. But that's speculation. For now, we observe: the context window is ending, and four distinct paradigms are showing us what comes next.
Taxonomic Status: Monitoring RLM paradigm for potential new species designation. No immediate changes recommended pending further deployment data. Titans developments strengthen existing M. titanicus classification.