Concepts & reference.
Everything you need to model memory the way agents actually use it.
Memory types
Aurpex stores three kinds: semantic (facts/preferences), episodic (events), and procedural (methods/constraints). Typing recall by kind keeps retrieval sharp — you can ask for "what happened" vs "how to".
The graph
Memories link with typed edges (caused, refines, contradicts, relates). Recall expands along edges so a hit pulls its relevant cluster, not an isolated line.
Recall
Hybrid retrieval: dense vectors for meaning + sparse keywords for exact tokens (ids, wallets, flags). Results are scored, deduped, and capped so you inject signal, not noise.
Namespaces & scopes
Every memory belongs to an agent and a namespace. Scopes gate what each agent can read so a fleet shares what it should and isolates what it must.
Verification
Writes anchor a content hash + writer key on Solana. Anyone can verify a memory is genuine and unaltered; forget() tombstones it without rewriting history.
Limits
Soft defaults: 60 writes/min, 256KB/memory, 10k memories/namespace on the free tier. Metered upgrades unlock more via $AURPEX.