Cold Start Problem (Cognitive)

In the AI Story

The contrast with warm context is instructive. A project that remains in working memory — because the builder engaged with it recently, or because its core elements are so well-learned they remain chronically accessible — requires minimal reactivation: a quick review of where things stood, a reminder of the next planned step, and the builder is operating near full capacity within minutes. A cold-start project requires extensive retrieval work that can take tens of minutes: reading previous notes to reconstruct decisions, reviewing AI outputs to remember what was tried, consulting specifications to re-derive goals, and mentally replaying the project's history to restore the causal understanding linking current state to past choices. The builder is nominally back on the project but operating with partial context for an extended period.

George Miller's classic finding that working memory holds 7±2 chunks suggests an upper bound on how many projects a builder can keep warm simultaneously. Each project, to remain accessible, must occupy at least one chunk; complex projects may require several. Assign a builder to three projects, and she might keep all three in working memory, allowing warm context switches between them. Assign her to twelve, and at least nine will be cold at any given time. The typical AI-augmented builder in 2026 is assigned to five to eight projects — a number that exceeds most individuals' working memory capacity and ensures that the majority of project returns will be cold starts.

The organizational cost is that most monitoring evaluations are performed with hastily reconstructed, incomplete context. The AI agent produces an output on Project 7, which the builder hasn't thought about in three days. She must cold-start: retrieve the project's goals, remember what was already tried, reconstruct the evaluative criteria, and make a judgment call — all in the minutes before the next output arrives on Project 3. The evaluation is performed with context that is both incomplete (some elements weren't retrieved) and corrupted by residue from the task she was working on when Project 7's output interrupted. The compounding of cold-start incompleteness and attention residue produces the worst possible conditions for judgment quality.

Workflow design can minimize cold starts by reducing the number of active projects per builder to the number that working memory can maintain. Three simultaneous projects allow warm switching; twelve guarantee cold starts. The organizational resistance to this recommendation is that assigning fewer projects per builder appears less efficient. But the efficiency calculation assumes that cognitive state doesn't affect output quality — an assumption Leroy's research directly refutes. The builder assigned to three projects, performing all evaluations with warm context and minimal residue, produces fewer total outputs but higher quality per output. The builder assigned to twelve, performing most evaluations with cold-reconstructed context and heavy residue, produces more total outputs but substantially lower quality per output. Which is actually more efficient depends on whether the organization values quantity or quality — and on whether it can measure the difference.

Origin

The cold start problem originates in computer science, where it describes the performance cost when a system must load resources from slow storage (disk) rather than fast storage (cache or RAM). The cognitive adaptation recognizes that human long-term memory is analogous to slow storage, working memory to fast storage, and that retrieving project context from long-term memory is orders of magnitude slower than refreshing context already in working memory. The term's application to AI-augmented multi-project work appears to have emerged from practitioners' lived experience of 'forgetting where I was' and needing to 'get back up to speed' before feeling competent to evaluate outputs on projects they hadn't touched in days.

Key Ideas

Full reconstruction required. When a project falls out of working memory entirely, return requires expensive retrieval from long-term memory rather than quick reactivation of already-accessible representations.

Error-prone process. Cold-start reconstruction is incomplete (some context is lost) and sometimes incorrect (retrieved context doesn't match actual project state), producing evaluations based on partial or corrupted understanding.

Guaranteed by over-assignment. Assigning builders to more projects than working memory can maintain (typically more than 3-5 complex projects) ensures most returns will be cold starts, maximizing cognitive costs.

Compounds with residue. Cold-start incompleteness combines with attention residue from the interrupted task to produce the worst possible evaluative conditions: partial context plus occupied working memory.