Read before you rely on it

Limits and honest scope

What AETHER-33 does not do, stated plainly. These are not roadmap promises.

Retrieval

• Lexical only. No embeddings, no paraphrase recall, no cross-lingual search.
• No reranking model, no query expansion, no synonym map.
• Ranking quality is BM25's quality; it is reproducible, not state of the art.

Safety

• Injection screening is pattern-based. Paraphrases, low-resource languages and semantic injections pass through.
• Contradiction detection is a lexical heuristic. It misses conflicts with no shared vocabulary and can flag nuanced updates.
• A quarantined memory is *stored*. The text exists until it is deleted; quarantine controls visibility, not persistence.
• Nothing here is a sandbox. The service does not execute the text it stores.

Operations

• One SQLite file per instance. No replication, no sharding, no automatic failover.
• Isolation is logical, not encryption at rest.
• No third-party security audit has been performed on AETHER-33.
• The per-tenant chain detects tampering by a party that does not also control your pinned heads. Pin heads outside this service if that matters.

Scale

The stress report on the measured-numbers page is a single-process run on this host with the corpus size stated. It is not a production capacity claim. If you need millions of memories, put a real index in front and keep AETHER-33 as the governed write log — or run the store on hardware you have measured yourself.