Q
What goes into a PackedContext?
Critical top-3 nodes (>75% score), ~5 patterns/bugfixes, ~5 concepts/approaches, plus token/time accounting — curated, not dumped.
Glossary · PackedContext
PackedContext.
2–4K tokens that beat 50K raw.
Direct answer
A PackedContext is a fixed-budget assembly of the best knowledge for one query: top critical nodes, best patterns and bugfixes, supporting concepts — selected submodularly (relevance minus redundancy) under a token cap, typically 2000/4096 across Knowledge/Thread/LowConf sections. It replaces two failure modes at once: whole-file dumps (expensive, noisy) and flat instruction files (stale, unbounded). Same task, ~85% fewer per-turn tokens, higher signal density.
In Anvaya
How We Implement It.
01PackedContextAnvaya's Quorum packer renders PackedContext as XML the model consumes directly; mind_query serves it over MCP under a 4096-token budget.
Questions
Asked About PackedContext.
Q
Why submodular selection?
Pure top-K retrieval returns near-duplicates. Maximal-marginal-relevance picks items that add new information given what's already packed.
Q
How is this different from RAG top-K?
RAG retrieves; packing budgets, diversifies, sections by confidence, and renders for the model's attention order.
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.