How to use serena memories in addition to files imported with claude md?

[nickolay-kondratyev]

My CLAUDE.md in the kotlin project looks as follows

@.ai_input/memory/0_project_overview.md
@.ai_input/memory/1_structure.md
@.ai_input/memory/2_build.md
@.ai_input/memory/2_test.md
@.ai_input/memory/3_kotlin_development_standards.md
@.ai_input/memory/4_testingStandards.md
@.ai_input/memory/4_testingStandards.noJsTarget.md
@.ai_input/memory/4_testingStandards.withJsTarget.md
@.ai_input/memory/5_js_target_support.md
@.ai_input/memory/6_version_management.md

When I ran onboarding from serena it generated memories that very much mimicked what was already in these files.

To me this raising a flag on maintaince of having duplicated instructions. So the question is there value keeping serena memories separate or should I symlink my .ai_input/memory/ directory into ./serena/memories so that serena has access to pre-existing "memory" (context input) files.

OR will even symlinking be just be duplicating the size of instructions in context, as claude.md is going to be loaded into the context, and instead we should keep serena memories much more concise? Trying to understand if there is something special with serena memories approach vs claude md files.

[opcode81]

Did Serena read the memories and create new ones based on their content or did it come up with similar contents on its own?
If it is the former, then adding instructions like this could help:

Memories must not duplicate existing content. If a memory's topic is alredy covered by another (topic-specific) file,
create a memory that only contains a reference to that file instead of duplicating content.

If you want to try this and it works well, feel free to provide a PR on that.

Symlinking could work also. The LLM is likely to not read an additional file on the same topic if it already knows enough about it.