Q
Does a 1M-token window fix context rot?
No — larger windows raise the ceiling but not attention quality. Rot is gradual, not a cliff; curation still wins at every size.
Glossary · Context Rot
Context Rot.
Why more context means worse answers.
Direct answer
Context rot is the measured degradation of model accuracy as input grows: every frontier model tested loses recall as windows fill, with semantic-matching tasks falling from 95%+ to 60–70% well before nominal limits. Attention is finite — each added token thins focus over everything already there, buries key facts mid-window (lost-in-the-middle), and raises cost on every subsequent turn. The defense is curation, not capacity: smallest high-signal token set per step.
In Anvaya
How We Implement It.
01Context RotAnvaya treats rot as the enemy: Prism + Quorum pack 2–4K tokens instead of ~50K raw, subagents isolate exploration, and output caps stop tool dumps from ever entering the window.
Questions
Asked About Context Rot.
Q
How do I detect rot in my sessions?
Contradictions of earlier decisions, vague short answers, forgotten loaded files. Behavioral signal beats token counters.
Q
What is lost-in-the-middle?
Models attend most to window start and end; facts buried mid-context get used least reliably (Liu et al., 2023). Front-load decisions, repeat critical constraints at both ends.
Stop Starting From Zero.
One binary. 11+9 Rust crates. 545 tests. Hand-written HNSW index. Three transport modes. Four providers, Ollama, Anthropic, OpenAI, Siemens. Zero API keys required to start. Mind remembers everything after the first session.