CONTEXDB
Exploring context storage queryable by structural shape rather than approximate string similarity.
THE BOTTLENECK
Vector databases index text embeddings or raw string chunks, relying on approximate nearest-neighbor search. For structured factual records and analytical queries, semantic vector similarity frequently fails on numeric filters, precise joins, and schema-level retrieval.
Every time an agent restarts a session, it re-reads and re-tokenizes the same underlying state.
THE THESIS
What does an inference-native context store look like when stored records are immutable compiled artifacts, and retrieval is indexed directly by structural schema rather than vector distance?
STORAGE SPECIFICATION
Investigated structural shape indexing and content-addressed storage architectures that allow agents to page relevant context into active prompts with microsecond latency.
SPECIFICATION DRAFT
Published initial specification draft (v0.1) defining structural schema contracts and content-addressed context blocks.
FINDINGS & PROGRESS
Established core architectural principles for separating active hot prompt state from warm compiled context storage.
NEXT STEPS
Benchmarking prototype storage paging latency against standard in-memory caches.