Sarah Wooders’ line, quoted in Harrison Chase’s article, captures the architectural claim precisely: “Asking to plug memory into an agent harness is like asking to plug driving into a car.” Short-term memory (conversation messages, large tool-call results) is managed by the harness. Long-term memory (cross-session state) must be written and read by the harness. How CLAUDE.md loads, how skill metadata is surfaced, what survives compaction, whether interactions are queryable — every one of these is a harness decision, not a decision a standalone memory service can make after the fact.
This has immediate bearing on Two-tier agent memory separates organizational workflow knowledge from individual user preferences and Persistent agent memory preserves institutional knowledge that walks out the door with employees: the separation between org and user memory is only meaningful if the harness exposes both tiers consistently to the model. It also complicates the popular “bring-your-own-memory” framing — until harness-memory abstractions standardize, memory isn’t portable across harnesses, only across backing stores. This suggests that choosing a harness is effectively choosing a memory model, which connects to Context is the product, not the model.