Foveance — Novelty & Positioning

This page keeps the project's claims honest. As of mid-2026 this space is crowded, and the broad idea is not new; the narrow, defensible core below is what Foveance actually claims.

What ALREADY EXISTS (cite, do not claim)

Prior work What it already does Why our headline can't be "this"
AFM — Adaptive Focus Memory (Cruz 2025, arXiv 2511.12712, code on GitHub) Per-message fidelity tiers FULL/COMPRESSED/PLACEHOLDER, scored by semantic-similarity-to-current-query + half-life recency + importance, packed under a token budget. This is the multi-fidelity-store-under-budget mechanism. Already done. We treat AFM as the primary baseline and as the "reactive" special case of our policy.
ContextBudget (2026) Context compression as a budget-constrained sequential decision problem, adapting to remaining window capacity. "Sequential + budget" is taken.
ACON (Kang et al. 2025, arXiv 2510.00615) Black-box, gradient-free compression of agent observations + history via failure-driven guideline optimization; API-applicable; distillable. "Black-box unified agent compression" is taken.
MemAct — Memory-as-Action (Zhang et al. 2025, arXiv 2510.12635) Context curation as RL policy actions (prune/write), MDP formulation, DCPO training. "Context management as MDP/RL" is taken (but requires training the agent; we don't).
Sequential Wyner–Ziv for KV cache (2026, arXiv 2605.25085) Rate–distortion limits of online cache compression as sequential Wyner–Ziv with next-step query as side info. RD/Wyner–Ziv framing exists — but white-box, next-step, KL distortion, single generation.
Nagle et al. (NeurIPS 2024, arXiv 2407.15504) Distortion–rate function for static black-box token-deletion prompt compression; shows query-awareness matters. RD for prompts exists — but one-shot, hard-deletion only.
RCR-Router, DAST, wireless-CE, Quest/TOVA importance-/relevance-aware budget or fidelity allocation in various single-shot or KV settings. "importance-aware allocation" is a known pattern.

What does NOT appear to exist anywhere (our defensible contribution)

The intersection is empty. Our four claims, each of which is individually checkable:

  1. Anticipatory (future-relevance) allocation criterion. Every prior system scores items by relevance to the current query (AFM), failure traces (ACON), or remaining budget (ContextBudget). We allocate fidelity by expected relevance to the future of the trajectory — an explicit posterior p(future needs | history) — and show it dominates the reactive criterion exactly when the task has cross-turn dependency structure (drift). The reactive policy is the drift = 0 special case.

  2. A fundamental-limits theory for the black-box, multi-turn, task-success setting. We define the trajectory distortion–rate function with task-success distortion and a future side-information variable (predictive Wyner–Ziv), generalizing Nagle (static → sequential) and distinct from the KV paper (white-box/next-step/KL → black-box/future/task). No existing work couples a fundamental-limits result to a deployable black-box agentic allocator.

  3. Two-sided successive refinement (re-inflation) with refinability conditions. We give conditions under which holding items at multiple fidelities is free (no rate penalty), licensing principled re-inflation. Prior systems compress one-way (AFM can re-upgrade reactively but offers no theory; ACON/summary methods are lossy/destructive).

  4. A near-optimal index policy with a measured greedy gap. We cast per-item fidelity as a multiple-choice knapsack / restless-bandit and use a Whittle/Lagrangian index that is provably within one item's value of the optimum, and we prove when local-greedy heuristics (AFM/Headroom/ContextBudget) are already near-optimal vs when anticipation is necessary. That "when is anticipation worth it" theorem is the practically useful, genuinely absent result.

One-sentence positioning

Prior work shows how to compress agent context under a budget; Foveance shows the fundamental limit of doing so over a trajectory, gives the anticipatory policy that approaches it, and proves when anticipation beats the reactive heuristics everyone currently ships.

Honesty guards for the paper

  • Lead the Related Work with AFM, ACON, ContextBudget, MemAct, the two RD papers. Position against them explicitly; reviewers in this area will know all of them.
  • The multi-fidelity store is substrate, not contribution. Say so.
  • Report the greedy gap honestly — if it's small on real tasks, that's a finding (it tells practitioners when cheap heuristics suffice), not a weakness.
  • Do not fabricate numbers. All result tables are filled from real runs (PROMPT_2).
  • If on real models reactive ≈ foveance, report it. The theory (when anticipation helps) must then explain why — that is still a publishable, honest contribution.