Ken W Alger published Part 2 of the "Building the AI Memory Stack" series on July 28, 2026, arguing that agentic AI systems need a layer called Active Working Memory between durable memory and the context window. He says the idea emerged while writing Part 1, when he moved among a glossary, "Memory as Infrastructure," Context Hydration notes, SDK specifications, Architecture Decision Records, GitHub, and an article draft. None of those sources contained the full answer alone; together, they formed a temporary working set needed for one task.

Alger compares the context window to a CPU cache: fast, temporary, and available only during the work being performed. He says caches do not populate themselves, and context windows do not either. In his model, Active Working Memory is the assembled execution state created before reasoning begins, after documents are retrieved, tools run, state is restored, policies are evaluated, and responses are normalized.

The article’s central distinction is “Models reason. Applications assemble.” Alger says a model does not retrieve documents, choose which Git commit matters, or know whether a tool response is stale; the application does those jobs. In a Python SDK update example, an agent retrieves an Architecture Decision Record describing a package boundary, loads the glossary definition of Active Working Memory, checks the current GitHub implementation, and only then builds the prompt.

Alger says many failures blamed on models are actually context assembly failures. The wrong evidence may have been retrieved, a stale document may have won, or a constraint may never have entered the working set. He frames those as architecture problems before they are model problems, because inference begins only after the application has decided what the model can and cannot see.

The proposed memory stack has 4 layers: Durable Memory preserves knowledge and is owned by storage or the application; Active Working Memory assembles task state and is owned by the orchestrator; the Context Window presents selected information and is owned by the runtime; the Model performs inference and is owned by the LLM. Alger says no layer substitutes for another, because a larger context window cannot fix poor selection and a better model cannot reason over evidence it never receives.

Alger separates search from context assembly. Search asks what information exists, while context assembly asks what information belongs together for a task. Two systems can use the same model, context window, and knowledge store but produce different results if one assembles a concise, relevant working set and another floods the model with loosely related material.

Treating Active Working Memory as a real layer turns context assembly from prompt engineering into systems architecture, according to Alger. Information crosses that boundary only after decisions about relevance, authority, recency, format, and priority. He says every unnecessary document adds Context Tax, every verbose tool response competes for attention, and every missing source creates a blind spot the model cannot detect from inside the context window.

The article closes by distinguishing temporary working sets from durable memory. Alger says the working set he assembled while writing disappeared after the article was finished, while the article remained. Agentic systems face the same choice: what belongs only in today’s working set and what should become tomorrow’s memory.

Ken W Alger published Part 2 of the "Building the AI Memory Stack" series on July 28, 2026, arguing that agentic AI systems need a layer called Active Working Memory between durable memory and the context window. He says the idea emerged while writing Part 1, when he moved among a glossary, "Memory as Infrastructure," Context Hydration notes, SDK specifications, Architecture Decision Records, GitHub, and an article draft. None of those sources contained the full answer alone; together, they formed a temporary working set needed for one task.

Alger compares the context window to a CPU cache: fast, temporary, and available only during the work being performed. He says caches do not populate themselves, and context windows do not either. In his model, Active Working Memory is the assembled execution state created before reasoning begins, after documents are retrieved, tools run, state is restored, policies are evaluated, and responses are normalized.

The article’s central distinction is “Models reason. Applications assemble.” Alger says a model does not retrieve documents, choose which Git commit matters, or know whether a tool response is stale; the application does those jobs. In a Python SDK update example, an agent retrieves an Architecture Decision Record describing a package boundary, loads the glossary definition of Active Working Memory, checks the current GitHub implementation, and only then builds the prompt.

Alger says many failures blamed on models are actually context assembly failures. The wrong evidence may have been retrieved, a stale document may have won, or a constraint may never have entered the working set. He frames those as architecture problems before they are model problems, because inference begins only after the application has decided what the model can and cannot see.

The proposed memory stack has 4 layers: Durable Memory preserves knowledge and is owned by storage or the application; Active Working Memory assembles task state and is owned by the orchestrator; the Context Window presents selected information and is owned by the runtime; the Model performs inference and is owned by the LLM. Alger says no layer substitutes for another, because a larger context window cannot fix poor selection and a better model cannot reason over evidence it never receives.

Alger separates search from context assembly. Search asks what information exists, while context assembly asks what information belongs together for a task. Two systems can use the same model, context window, and knowledge store but produce different results if one assembles a concise, relevant working set and another floods the model with loosely related material.

Treating Active Working Memory as a real layer turns context assembly from prompt engineering into systems architecture, according to Alger. Information crosses that boundary only after decisions about relevance, authority, recency, format, and priority. He says every unnecessary document adds Context Tax, every verbose tool response competes for attention, and every missing source creates a blind spot the model cannot detect from inside the context window.

The article closes by distinguishing temporary working sets from durable memory. Alger says the working set he assembled while writing disappeared after the article was finished, while the article remained. Agentic systems face the same choice: what belongs only in today’s working set and what should become tomorrow’s memory.