A long-running project contains more than source files. It also contains decisions, conventions, learned workflows, useful relationships and evidence about what has already worked. Losing that context between AI sessions creates repeated discovery work.
Durable does not mean unbounded
A project memory layer is most useful when it stores the smallest durable representation that can help future work. Raw secrets, unrestricted terminal history and machine-specific credentials do not need to become long-term AI context.
- Keep decisions and project rules that affect future implementation.
- Preserve verified experience with provenance instead of promoting guesses to fact.
- Retrieve only the context that is relevant to the current task.