Claude Code found my app’s shared contract in two MCP calls. Then the records ran out.

Claude Code users often start a new session with no prior knowledge of the system they are working on. The agent must rediscover the architecture, including which services and frontends interact and what data types they share. This process often involves memory, as Claude Code loads CLAUDE.md instructions and its own auto-memory notes to recall important information from previous sessions.

However, in September 2023, Claude Code sessions using a redesigned Projects feature gained the ability to share memory across threads. This allowed Claude to access a shared memory of records that provide context without relying solely on memory.

To test this alternative source of context, I built a small support console using Claude Code and examined how it could recover context from dependency records alone. The console consisted of a React app, an Express service backed by MongoDB, a shared ticket entity, and a platform component that connects everything. All four components were versioned publicly in the scope "bit-oss.support", and the console was deployed in production over HTTPS.

When I asked Claude to add filtering by ticket status and assignee, it first checked the empty workspace for local components. Since none existed, Claude read the scope "bit-oss.support" through MCP and found the four components and their dependencies. It then read the API references and file inventories for the app, service, and ticket entity, importing their source without any prior knowledge of the project.

Claude discovered that the app and service depended on the same version of the ticket entity, which was described as "shared between the support service and the agent-facing app." The API reference confirmed this shared dependency.

Claude imported all four components using `bit import`, inspected the source code, and implemented client-side filtering within the app. Tests showed 41 passing tests, including nine new filtering tests. Claude used MPC to discover the system structure, and source inspection and tests validated how to modify the system without affecting the shared ticket entity or the service.

The dependency records had limitations, running out at behavior. While the records revealed which components shared the ticket contract, they did not check if the routes on each side still matched the real service. Additional engineering work was required to ensure proper behavior, including concurrency control, migrations, recovery procedures, and mandatory database tests.

Memory files, such as CLAUDE.md instructions and auto-memory notes, hold valuable intent and context written by humans. Dependency records, on the other hand, provide structure and meaning, showing exactly which components depend on which versions, without any drift from the described versions. Together, memory and records offer a comprehensive understanding of a system's context and functionality.