{"slug": "active-working-memory-the-ram-of-agentic-systems", "title": "Active Working Memory: The RAM of Agentic Systems", "summary": "Ken Walger, in the second part of his Building the AI Memory Stack series, introduces the concept of Active Working Memory as the second layer in an AI memory hierarchy, arguing that the context window is merely the execution surface. He explains that agentic systems assemble a working set of retrieved documents, tool outputs, and state before the model begins reasoning, and that failures blamed on the model are often failures of context assembly.", "body_md": "*Part 2 of the Building the AI Memory Stack series*\n\nWhen I published the [first article](https://www.kenwalger.com/blog/ai-engineering/architecture/context-window-is-not-memory/?utm_source=devto&utm_medium=article&utm_campaign=building-the-ai-memory-stack) in this series, I thought I was writing about context windows.\n\nThe more I wrote, the more I found myself bouncing between documents. I had the glossary open in one browser tab. The [Memory as Infrastructure](https://sovereignplatform.dev/terms/memory-as-infrastructure.html?utm_source=devto&utm_medium=article&utm_campaign=building-the-ai-memory-stack) article was open in another. A third tab contained notes about [Context Hydration](https://sovereignplatform.dev/terms/context-hydration.html?utm_source=devto&utm_medium=article&utm_campaign=building-the-ai-memory-stack). GitHub was open with the SDK specifications, and a handful of Architecture Decision Records were sitting beside my editor.\n\nNone of those documents individually contained \"the answer.\" Together, they formed the temporary collection of information I needed before I could make progress.\n\nThen something unexpected happened.\n\nI was no longer writing about context windows.\n\nI was reconstructing a memory hierarchy, and the context window was only its first layer.\n\nThat is why this became a series.\n\nIn Part 1, I argued that the context window is best understood as the CPU cache of an AI system. It is an execution surface, not a memory system. Like CPU cache, it is optimized for fast access and exists only for the duration of the work being performed.\n\nBut caches do not populate themselves.\n\nNeither do context windows.\n\nBefore I could write, I assembled a working set.\n\n```\nBrowser tabs and open documents\n\n- Sovereign Systems glossary\n- Memory as Infrastructure\n- Context Hydration notes\n- SDK specifications\n- Architecture Decision Records\n- GitHub repository\n- Article draft\n\n            |\n            v\n\n    Active Working Memory\n\n            |\n            v\n\n      Context Window\n\n            |\n            v\n\n         Reasoning\n```\n\nThat collection was not my long-term memory. It was a temporary working set assembled for one task. My brain still had to compare ideas, notice contradictions, and produce something new. The documents simply gave the reasoning process the information it needed.\n\nAgentic systems work in much the same way.\n\nBefore a model begins reasoning, documents have been retrieved, tools have executed, state has been restored, policies evaluated, and responses normalized. The prompt is usually the final artifact produced by an orchestration layer, not the beginning of one.\n\nBy the time the model receives its first token, dozens of retrieval, filtering, ranking, and assembly decisions may already have been made.\n\nThat assembled execution state is what I call ** Active Working Memory**.\n\nThis article is about the second layer in that stack.\n\nOne sentence captures the distinction this entire article is trying to make.\n\nModels reason. Applications assemble.\n\nThose are different responsibilities.\n\nA model does not retrieve documents.\n\nA model does not decide which Git commit matters.\n\nA model does not know whether a tool response is stale.\n\nAn application does.\n\nConsider a concrete case. An agent is asked to update a Python SDK.\n\nIt retrieves the ADR describing the package boundary.\n\nIt loads the glossary definition of Active Working Memory.\n\nIt checks the current implementation on GitHub.\n\nOnly then does it build the prompt.\n\nPeople often talk about \"putting something into the context window.\"\n\nThat wording quietly suggests the context window is responsible for finding, selecting, and organizing information.\n\nIt isn't.\n\nBy the time inference begins, the application has already decided what the model will, and will not, be allowed to see.\n\nThe context window does not begin the process. It receives the result of the process.\n\nMany failures blamed on the model are actually failures of context assembly. The wrong evidence was retrieved. A stale document won. A constraint never made it into the working set. Those are architecture problems before they are model problems.\n\nIf the context window is the CPU cache, Active Working Memory is the system RAM.\n\n| Layer | Primary Responsibility | Typical Owner |\n|---|---|---|\n| Durable Memory | Preserve knowledge | Storage / Application |\n| Active Working Memory | Assemble task state | Orchestrator |\n| Context Window | Present selected information | Runtime |\n| Model | Perform inference | LLM |\n\nEach layer has a distinct responsibility. No layer substitutes for another. A larger context window does not repair poor selection, and a better model cannot reason over evidence it never receives.\n\nSearch answers one question.\n\n*What information exists?*\n\nContext assembly answers another.\n\n*Given this task, what information belongs together?*\n\nThose sound similar. Architecturally, they are completely different.\n\nActive Working Memory presents that second decision to the model.\n\nTwo systems can use the same model, the same context window, and the same knowledge store yet produce very different results, because one assembles a concise, relevant working set while the other floods the model with loosely related information.\n\nThe model is identical.\n\nThe memory architecture is not.\n\nOnce Active Working Memory is treated as a real layer, context assembly stops looking like prompt engineering and starts looking like systems architecture.\n\nInformation crosses this boundary only after decisions have been made about relevance, authority, recency, format, and priority.\n\nEvery unnecessary document increases [Context Tax](https://sovereignplatform.dev/terms/context-tax.html?utm_source=devto&utm_medium=article&utm_campaign=building-the-ai-memory-stack). Every verbose tool response competes for attention. Every missing source creates a blind spot the model cannot recognize from inside the window.\n\nThe context window can only reason over what Active Working Memory hands it.\n\nThe working set I assembled while writing this article disappeared as soon as the article was finished.\n\nThe article remained.\n\nThat distinction turns out to matter.\n\nAgentic systems face the same decision. What belongs only in today's working set? What deserves to become tomorrow's memory?\n\nThat is where Part 3 begins.\n\nModels don't assemble context. They inherit it.\n\n**Working memory is not where knowledge lives. It is where knowledge collaborates.**", "url": "https://wpnews.pro/news/active-working-memory-the-ram-of-agentic-systems", "canonical_source": "https://dev.to/kenwalger/active-working-memory-the-ram-of-agentic-systems-4ldh", "published_at": "2026-07-28 18:21:00+00:00", "updated_at": "2026-07-28 18:35:45.974983+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-agents", "ai-infrastructure"], "entities": ["Ken Walger", "Sovereign Systems", "GitHub"], "alternates": {"html": "https://wpnews.pro/news/active-working-memory-the-ram-of-agentic-systems", "markdown": "https://wpnews.pro/news/active-working-memory-the-ram-of-agentic-systems.md", "text": "https://wpnews.pro/news/active-working-memory-the-ram-of-agentic-systems.txt", "jsonld": "https://wpnews.pro/news/active-working-memory-the-ram-of-agentic-systems.jsonld"}}