Skip to main content

What NOT to Save

What: Not everything belongs in memory. Saving too much creates noise that dilutes the signal of genuinely important persistent knowledge. Memory should contain judgment calls and context — things the AI cannot derive from the code itself.

Do not save:

CategoryWhy NotWhere It Lives Instead
Code patternsReadable from the source code directly. Memory about code drifts as code changes.The code itself. Reference memory can document why a pattern exists but should not duplicate the implementation.
Git historyAccessible via git log. Duplicating commit history in memory wastes context.Git.
Debugging session detailsEphemeral. The specific port conflict or import error from Tuesday's debugging session has no lasting value.Nowhere — let it go.
Test commandsDocumented in CLAUDE.md (project instructions).CLAUDE.md's testing section.
File-level documentationBelongs in code comments, docstrings, or README files co-located with the code.The source files themselves.
Current task statusChanges too rapidly. By the time memory is read in the next session, the status is stale.Conversation context within the current session.

The filtering principle: Before writing a memory, ask: "If Claude starts a fresh session tomorrow and does NOT have this information, what will go wrong?" If the answer is "nothing — it can figure this out from the code or CLAUDE.md," do not create a memory file. If the answer is "it will repeat a mistake we already corrected" or "it will propose an approach we already rejected," create the memory.

Evidence: Trust Relay's memory system has 27 files, not 270. The memory system is selective. It captures decisions (why we chose Temporal over Celery), preferences (always use Alembic), corrections (never blanket-delete MinIO), and state (which pillars are complete). It does not capture code snippets, debugging logs, or test output. The result is a memory corpus that Claude can read in full at session start without consuming excessive context window — the hub file (MEMORY.md) stays inside its 24,000-byte budget because it links to topic files rather than inlining everything.