Claude Code and custom repo memory solve different problems, developer test finds

A developer on DEV Community ran a controlled test comparing Claude Code's built-in session memory against a custom repository-bound memory system across multiple editors, machines, and users. Both systems successfully recalled a user-defined fact within the same session and editor, but diverged when the query came from a different tool or colleague. The key distinction found was that vendor memory keeps a single assistant consistent with itself, while repository memory keeps project knowledge accessible across people, machines, and model upgrades. The developer also identified a structural limitation of vector similarity search: it cannot reliably link causally related entries, such as an error log and its fix discovered weeks later, because semantic closeness and causal connection are different relationships. The author acknowledged the findings cut against their own product, noting that for simple retrieval needs, a plain vector search is often sufficient without additional tooling.