Volume II · Context is the finite resource. Engineering it is the discipline.

The Science of Context

Context windows are finite, and how a session fills them decides whether an agent stays sharp or drifts. This volume traces the arc from Karpathy's context engineering coinage through context rot, health scoring, compaction contracts, and the memory ladder — from static retrieval to agent-maintained memory — closing with the full context lifecycle: created, spent, lost, recovered, persisted.

Chapters

Published

01Why Context Beats PromptsKarpathy's coinage — 'the delicate art and science of filling the context window with just the right information for the next step' — to Anthropic's formalization of context as a critical-but-finite resource.7 min readPublished

02Context RotThe Chroma Research term: models don't use context uniformly, and reliability degrades as input grows — why more context is not better context.8 min readPublished

03Context HealthA measurable health model — score, band, named reasons, one-click remediations.7 min readPublished

04Compaction ContractsWhat survives summarization, and designing for it: the survival-contract pattern of what gets cleared, summarized, or lost.7 min readPublished

05Context RecoveryRe-grounding after loss: session continuity, handoff summaries, restart-with-handoff.7 min readPublished

06Context DriftDivergence between what the agent believes and what's true — files edited outside the session, stale references, the rot-by-external-edit problem.9 min readPublished

07The Memory LadderRAG's static retrieval gives way to agentic search and then agent memory, where the agent also writes to the store.7 min readPublished

08File-Based MemoryThe converged channels — git history, progress logs, persistent task state, AGENTS.md/CLAUDE.md — that let learned patterns carry across stateless iterations.8 min readPublished

09Decision MemoryDecisions as first-class, queryable context: what was decided, why, and what it superseded — the "why did it do that?" query.7 min readPublished

10Knowledge Graphs as ContextStructured context: entities, relations, and when graph beats prose.7 min readPublished

11The Context LifecycleCapstone — the context lifecycle: created, spent, lost, recovered, persisted.5 min readPublished

Key sources: Anthropic "Effective context engineering for AI agents" · Chroma "Context Rot" · Anthropic "Effective harnesses for long-running agents" · Awesome-Context-Engineering survey repo (collision check) · RAG→memory ladder literature

Related