Context Flow

Context is a pipeline, not a pile. An artifact is admitted into an actor's window, resident for a span, then evicted. During residency it has a relative placement within the window and a changing recall profile: recall distance is the proportion of resident context admitted between the artifact's admission and an attempted use; recall efficacy is percentage accuracy on those attempts. All three belong to the resident artifact — the artifact, not the signal it carries, is the unit of recall.

The context-engineering family has prices (the economy) and moves (the mechanisms), but without flow it has no model of movement. Context gets discussed as a pile — a window that is some percentage full — while the questions that shape an architecture have no vocabulary: what enters whose window, when, where it sits, how far it must travel before use, how accurately it can be recalled, and when it leaves.

• Model — the lifecycle verbs. Admission is the event that loads an artifact into a window. Its trigger may be the actor's own boot: inherited context such as the system prompt or memory-file lines arrives by lineage, chosen by nobody downstream. Residency is the span the artifact stays, and what statelessness re-bills every turn. Eviction ends residency through compaction, summarization, truncation, a dispatch returning, or a session ending. Eviction ends recall from the artifact and removes its steering trajectory (Momentum's second motive for eviction, beyond the budget arithmetic).

• Model plus empirical result — recall during residency. In an append-growing window, recall distance can change an artifact's placement. Controlled evaluations show that recall distance and placement can each change recall efficacy — and the wider resident context matters on its own. The direct distance-to-efficacy path captures separation and interference that can matter even without a coarse placement change. The resulting profile need not decline monotonically and is conditional, not universal (Liu et al. 2024; Veseli et al. 2025; Du et al. 2025).

  recall distance ──▶ placement ────────┐
         │                              ▼
         └──────────────────────▶ recall efficacy ◀── wider resident context
  

• Measurement model — state, flow, outcome. Placement is the artifact's state at the moment of recall, not merely where it was admitted. Describe it proportionally relative to both window boundaries rather than with durable token thresholds. The two boundary readings are complementary views of the same relative placement; total resident length remains a separate covariate. Recall distance is the flow of intervening context since admission, not wall-clock delay. Under ordinary appending it changes placement; reconstruction or reordering can make the two diverge. Recall efficacy is the outcome: one attempt produces a scored result, while comparable attempts yield percentage accuracy. Within this model efficacy does not feed back into the placement or distance already observed.

• Model — artifacts carry signals. Placement, recall distance, and recall efficacy belong to the artifact actually resident. A signal is the fact, instruction, or evidence that may continue across representations. Summarizing an artifact evicts the source and admits a new artifact — even when the new one preserves the same signal or a subset. The replacement starts its own placement and recall-distance history. Information omitted by the transformation is a transformation-fidelity failure, not a recall failure against the replacement.

Placement is a property during residency, not a lifecycle stage. Recall is an attempted use during residency, not a fourth stage. One artifact may therefore have several recall measurements over one residency.

Recall efficacy is behavioral accuracy, not the artifact's raw attention weight, salience, or influence. Those may help explain a positional effect, but they are not substitutes for the measured outcome.

Decay can reduce recall efficacy while an artifact remains resident — but decay is not eviction, the event that ends recall from the artifact.

This is a concept — the model, not a move. It does not prescribe when to admit, what to evict, whose window to use, or where to place an artifact. Those are objective — value per unit of context spent (cost / state / outcome lenses)'s questions. Nor does it claim a universal positional curve or fixed threshold; each profile is conditional on the model, task, prompt, artifact, resident context, and scorer.

• The mechanisms become comparable moves. Deferral delays admission and shortens residency; distillation shrinks what is admitted; subagent offload routes a segment through another actor's window; a dispatch return evicts the whole working set at the boundary.
• Eviction has two consequences. It ends the economy's residency bill and removes the trajectory that momentum says resident context exerts.
• A transformation restarts the artifact lifecycle. A summary may carry forward all or part of the source's signal, but its admission begins a new placement and recall-distance history.
• Recall claims need their conditions. Percentage accuracy means little without the artifact, recall distance, placement, wider resident context, model, task, prompt, and scorer under which it was measured.
• A residency can be sampled, not summarized by one score. Repeated recall attempts observe the same artifact at different placements and recall distances, producing a profile over its residency rather than one permanent efficacy value.

• Placement can change accuracy, but no curve is universal. Moving the same answer-bearing material within a fixed prompt changes accuracy in many controlled long-context evaluations, while some models and tasks remain robust. Primacy can weaken or reverse as relative window occupancy grows (Liu et al. 2024; Veseli et al. 2025).
• Distance and wider context can matter separately. Intervening material can reduce accuracy through separation or interference, while total input length can hurt even when relevant evidence remains immediately before the question (Wang et al. 2025; Du et al. 2025).
• Attention is not efficacy. Positional-attention studies offer possible mechanism evidence, but the recall measure remains scored output accuracy (Hsieh et al. 2024).
• Lifecycle is the model's inference over the evidence. Most controlled studies construct a completed prompt and vary position or intervening material. Mapping those results onto an artifact moving through an append-growing window is structural inference, not a directly observed multi-turn lifecycle. The full research trail is in the placement-and-recall synthesis.

• Dynamics, beside statics and objective: The Context Economy prices what residency costs; this node models how context moves; objective — value per unit of context spent (cost / state / outcome lenses) says what to optimize.
• The mechanisms execute on this pipeline: Deferred Context, Subagent Offload, and the travelling declared-shape contract specialization are each a move described in this node's verbs.
• Child: Momentum — a dynamics phenomenon of the pipeline: resident context does not just inform the model, it steers it.
• Lifecycle and recall-term definitions (admission, residency, eviction, placement, recall distance, recall efficacy) live in the glossary; this node teaches the model.

The working set model (Denning, CACM, 1968) and stack distance (Mattson et al., Evaluation Techniques for Storage Hierarchies, 1970): admission, residency, and eviction under a capacity bound, with reuse counted in intervening references, not elapsed time. Recall distance is that measure over admitted context. The delta: a cache hit is binary and placement-blind, while residency guarantees nothing and placement changes the odds.

• Live-lifecycle evaluations: most evidence manipulates completed prompts, not repeated recall attempts across one artifact's residency.
• Flavor-lane candidate: the pipeline metaphor is inherently narratable (admission gates, residency meters, eviction valves).