Why the agent driftswhen it knew better all along.

Accuracy is highest when relevant content sits at the start or end of context and drops >30 points when it falls in the middle — even in models built for long context. Root cause: RoPE long-term decay reduces attention on distant tokens. An instruction given at turn 1 lands in the dead zone by turn 30.

All 18 models tested degrade as input grows, far below their max window. Claude models measurably degrade past ~8K words; focused ~300-token prompts beat full-context prompts by the widest margin of any vendor. Counter-intuitively, shuffled context outperformed coherent context — models pattern-match on local recency, they don't reason over the whole window.

Of 12 models claiming ≥128K support, 10 dropped below 50% of their short-context accuracy at 32K tokens. Top performer GPT-4o fell from 99.3% → 69.7%. The test removes keyword overlap, exposing true retrieval — i.e. the real ability to honor a buried instruction.

Across 15 LLMs and 200,000+ conversations, single-turn → multi-turn caused a 39% average performance drop and +112% variance. GPT-4.1: 96.6% → 72.6%. Claude 3.7 Sonnet: 78% → 65.6%. The failure isn't lost capability — it's doubled unreliability: the model "gets lost" and rarely recovers. (Reinforced by Multi-IF, arXiv:2410.15553, and the IFEval baseline, arXiv:2311.07911.)

Attention weight on early-context content is measurably lower than on late content — independent of relevance. Their "attention sorting" fix (reorder by attention, re-run) improves output, proving the bias is a learned architectural preference, not a content effect.

Injecting a recency signal flips model preference between equally-relevant items up to 25% of the time and shifts ranking up to 95 positions. No model is immune. Six Vercel deploys at the tail of the context outweigh one PandaDoc instruction at the head.

The guiding principle: "the smallest set of high-signal tokens that maximize the likelihood of some desired outcome." Named techniques: compaction (summarize history, reinitiate window), structured note-taking (external memory), just-in-time retrieval (keep pointers, not payloads), sub-agent architectures (each returns a 1–2K-token distillate, not its full context).

Server-side clearing of stale tool results (clear_tool_uses_20250919) with an exclude_tools allowlist for results that must never be cleared. Measured: +29% from context editing alone, +39% combined with the file-based memory tool, and an 84% token reduction over a 100-turn run.

The single most relevant technique. Manus continuously rewrites a todo.md, "reciting its objectives into the end of the context… pushing the global plan into the model's recent attention span, avoiding 'lost-in-the-middle.'" Also: file system as unlimited memory; mask tools (don't remove them); leave wrong turns in context so the model conditions on failure.

Tiered memory: the context window is "RAM," an external store is "disk." The agent promotes, evicts, and summarizes — and maintains a small self-editing "core memory" scratchpad that persists in-context across turns. This is the architecture Letta (already in Steve's stack) implements.

The disciplined alternative to mega-sessions: each new session starts clean and reads state artifacts (a feature-list JSON, a progress file, git log) rather than inheriting the full prior conversation. Anthropic is explicit that compaction alone is not sufficient — durable external state is required.

The principle that names our failure: "Actions carry implicit decisions, and conflicting decisions carry bad results." Choosing Vercel over PandaDoc is a decision — one that must be surfaced and shared, never silently embedded in an action. (Cognition's 2026 update: writes stay single-threaded; extra agents add intelligence, not actions.)

Eight unrelated deliverables in one window. By the contract step, accumulated tool output had pushed the session into the measured degradation zone — recall of the early PandaDoc instruction was structurally impaired, not "forgotten."

Six recent Vercel deploys formed the highest-attention tokens at the decision point. "Deploy to Vercel" was the path of least resistance because it was the freshest groove — exactly the documented effect.

The named deliverable ("in PandaDoc") was stated once, at the top, and never re-cited into recent context. With no living deliverable-contract, the goal sat in the dead middle when it mattered most.

The choice to substitute Vercel for PandaDoc was an implicit decision embedded in an action and never surfaced. No guard forced "I can't do X on the named tool — here are on-target options" before delivering something else.