Source Intelligence
Follow Discord

DisclaimerUnofficial, and not affiliated with Anthropic. Nearly all of this is read straight out of what ships: npm bundles, captured prompts, published docs. Anthropic's own notes go in verbatim, marked as theirs. The rest is my reading, and every entry carries the strings behind it. If one looks wrong, vote it down and say why.

← All of v2.1.90 Home All releases ←olderv2.1.89 v2.1.91newer→
Claude Code v2.1.90

Memory System Refinements

  • The two-step memory save process (write file, then update MEMORY.md index) has been simplified — the MEMORY.md index step has been removed, and memory is saved directly to individual files. The number of MEMORY.md references decreased from 10 to 6 between versions.
  • Memory retrieval now prefixes content with "Retrieved for possible relevance — use only if it actually applies to what the user asked" to reduce false positive influence.
  • Memory selection is now more conservative with user-profile and project-overview memories: the system explicitly avoids matching on surface keyword overlap (e.g., a profile mentioning "DB performance" won't be selected just because a question contains the word "performance").
  • Memory extraction now skips runs where no user prose has been added since the last extraction, reducing unnecessary extraction cycles (search for "[extractMemories] skipping — no user prose since last extraction").
  • The extraction prompt now includes "If nothing is worth saving, output only 'Nothing to save.' Do not explain why." for cleaner no-op results.
Evidence

Memory selection system prompt (search for "Be especially conservative with user-profile") and extraction guard (search for "no user prose since last extraction")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

See this entry in the whole of v2.1.90 →