Relationship memory
Relationship memory is a structured record of the people you know, held outside any one app or AI assistant, so whatever tool you use next can read it.
Relationship memory is the layer underneath a personal CRM: not the interface, but the record itself. People, organizations, events, the typed links between them, and a source behind every fact. The distinguishing property is that it is not owned by the thing reading it. An assistant's built-in memory ends at that assistant; a CRM's database is only reachable through that CRM. Relationship memory is designed to be portable and read by several clients at once.
The idea became practical in 2025 and 2026, when the Model Context Protocol gave assistants a standard way to read an external system. The consequence follows directly. If several assistants can read the same store, the store outlives any one of them, and the value sits with whichever layer accumulates rather than whichever layer reasons. Models are released on a months-long cycle and are increasingly interchangeable. A twenty-year record of who you know is neither.
Relationship memory is a specific case of a general distinction worth keeping straight: conversational memory stores what was said to an agent, and relationship memory stores who somebody knows. The first is a transcript problem and is well served by generic memory infrastructure. The second is a schema problem, because a question like which of my contacts work in venture capital is a join over typed entities, not a similarity search over remembered sentences.
Why it matters
You will change AI assistants several times this decade, and most people will run more than one at a time. Anything an assistant remembered about your network leaves with it unless the memory was deliberately stored elsewhere. Relationship memory is what makes twenty years of accumulated context survive the tool churn underneath it.