Make the boundary explicit
Your application owns its user experience and workflow. The model produces its responses. thinqOS can maintain a Mind and expose capabilities through MCP and API interfaces. Your integration decides when those capabilities are used.
Saving context does not make it appear in every response automatically. Capture, retrieval, identity and prompt assembly must all be connected. Test each boundary with one fact you can trace.
A small integration you can verify
Save a harmless project preference for a test identity. Start a fresh session. Ask for relevant Mind context before the model answers, then inspect whether that preference was included and used appropriately. Correct it and repeat.
Keep durable instructions separate from inferred lessons. Treat retrieved material as evidence with a source, not a higher-priority instruction. A document or outside message must not acquire authority just because it was remembered.
Use both kinds of retrieval
A question about a general preference calls for different evidence than a request for an exact invoice number. Semantic search finds related meaning. Lexical search finds words and identifiers. A useful integration preserves access to original records as well as selected Mind context.
Measure misses, contradictions, latency and total cost on your own tasks. A retrieval function returning successfully is not evidence that it found the needed fact.
Keep an exit path
Mind exports support a signed archive and open representations including JSON-LD, YAML-LD and a Markdown vault. Verify which layers and fields your use requires.
Portable data is not a portable running system. Tool connections, application behavior and the cognition implementation require their own migration plan. Start with the setup guide and inspect the export before depending on it.