[wzxsdf]

Description

The feedback update guard used character-set similarity to decide whether a memory update was small enough to keep.

This caused reordered multi-fact memories to be treated as minor changes because the old and new texts contained similar character sets. A destructive rewrite could therefore pass the update guard.

This change uses difflib.SequenceMatcher, which considers character order as well as content. A regression test covers reordered text with the same facts.

Related Issue: #2527

No new dependencies are required.

Type of change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality not to work as expected)
• Refactor (does not change functionality, e.g. code style improvements)
• Documentation update

How Has This Been Tested?

• Unit Test
• Test Script Or Test Steps
• Pipeline Automated API Test

Validation performed:

• poetry run pytest tests/mem_feedback -q — 21 passed
• Ruff check passed
• Ruff format check passed

Checklist

• I have performed a self-review of my own code
• I have added tests that prove my fix is effective
• I have linked the issue to this PR
• I have mentioned the person who will review this PR

Reviewer Checklist

• closes Correcting or restoring one fact in a multi-fact memory rewrites the entire memory text #2527
• Made sure Checks passed
• Tests have been provided

[Memtensor-AI]

🤖 Open Code Review

Target: PR #2544
Task: 26f91287863d7d24
Base: main
Head: codex/fix-feedback-update-similarity

🔍 OpenCodeReview found 1 issue(s) in this PR.

---

1. src/memos/mem_feedback/utils.py (L47)

SequenceMatcher defaults to autojunk=True, which for sequences of 200+ elements treats any element occurring more than len/100 + 1 times as "popular" and removes it from matching. Memory texts of a few hundred characters are common here, and in them spaces and frequent letters/CJK characters exceed that count, so ratio() under-reports similarity even for near-identical texts (matching blocks break at every popular character). The same applies to the word-level SequenceMatcher below once a text has 200+ words.

Effect on this guard: change_ratio is inflated, so for old_len >= 200 (the 0.2 threshold branch) ordinary small edits — and even identical texts — get should_keep_update() == False and are downgraded from UPDATE to ADD in standard_operations, creating duplicate memories instead of updating the existing one.

Pass autojunk=False explicitly (and be aware ratio() is O(n*m), so very long texts get more expensive than the previous O(n) set check).

💡 Suggested Change

Before:

        character_similarity = SequenceMatcher(None, text1, text2).ratio()

After:

        character_similarity = SequenceMatcher(None, text1, text2, autojunk=False).ratio()

Generated by cloud-assistant via Open Code Review.

[Memtensor-AI]

✅ Automated Test Results: PASSED

All tests passed (11/11 executed). memos_python_core/changed-repo-python: 11/11. Duration: 18s

Branch: codex/fix-feedback-update-similarity