fix: prevent memory drift for referential queries in fast mode

[Zhe-SH-CN]

Description

In fast search mode, TaskGoalParser._parse_fast ignored conversation history entirely. Referential queries like "再找找其他价格优惠的" were embedded as-is, causing them to match unrelated older memories instead of the current conversation topic (issue #1365).

This PR adds lightweight context-aware query expansion in fast mode:

• When conversation history is available and the query looks referential (short, or contains referential hints like 其他/这个/再/more/another), prepend the most recent user message to the query before embedding
• This is a zero-LLM-call heuristic — no latency added to fast mode
• Long user messages are truncated to 120 chars to avoid over-long embedding inputs
• Self-contained long queries (> 20 chars, no referential hints) are not modified

Related Issue: Fixes #1365

Type of change

• Bug fix (non-breaking change which fixes an issue)

How Has This Been Tested

• Unit tests covering: no-conversation, referential query augmentation, non-referential query passthrough, English referential words, empty conversation, conversation without user messages, long message truncation, and public parse() method integration
• python3 -c "import ast; ast.parse(...)" syntax verification

Checklist

• I have performed a self-review of my own code
• I have commented my code in hard-to-understand areas
• I have added tests that prove my fix is effective or that my feature works
• I have linked the issue to this PR

[Memtensor-AI]

🤖 Open Code Review

Target: PR #2154
Task: 56119b2d819611c1
Base: dev-v2.0.24
Head: fix/query-context-drift

🔍 OpenCodeReview found 6 issue(s) in this PR.

---

1. tests/memories/textual/test_task_goal_parser_context.py (L83-L86)

The truncation assertion is vacuously weak. "A" * 120 is a substring of the full 300-character string, so assert "A" * 120 in result.rephrased_query would pass even if truncation is never applied (i.e., if the full 300 'A's were kept). While len(result.rephrased_query) < 300 partially closes the gap (the un-truncated form would be ~304 chars), it does not directly assert the key invariant. Add a strong negative assertion to make the test meaningful:

assert "A" * 121 not in result.rephrased_query

💡 Suggested Change

Before:

        result = parser._parse_fast("再找找", conversation=conversation, context="")
        # The augmented query should contain the truncated version (last 120 chars)
        assert "A" * 120 in result.rephrased_query
        assert len(result.rephrased_query) < 300

After:

        result = parser._parse_fast("再找找", conversation=conversation, context="")
        # The augmented query should contain the truncated version (last 120 chars)
        assert "A" * 120 in result.rephrased_query
        assert "A" * 121 not in result.rephrased_query  # Verify truncation actually occurred

---

2. tests/memories/textual/test_task_goal_parser_context.py (L61-L62)

The test only verifies that the prepended context appears in rephrased_query, but never asserts that the original query ("show me more") is also preserved. If the implementation accidentally dropped the query and only kept the context prefix, this test would still pass. Add an assertion for the original query:

assert "show me more" in result.rephrased_query

💡 Suggested Change

Before:

        result = parser._parse_fast("show me more", conversation=conversation, context="")
        assert "Find me a red dress" in result.rephrased_query

After:

        result = parser._parse_fast("show me more", conversation=conversation, context="")
        assert "Find me a red dress" in result.rephrased_query
        assert "show me more" in result.rephrased_query

---

3. src/memos/memories/textual/tree_text_memory/retrieve/task_goal_parser.py (L98-L106)

False-positive referential detection due to single-character substring matching.

Several hints in _REFERENTIAL_HINTS are single Chinese characters ('也', '还', '再', '他', '她', '它'). Because the check is a plain substring test (h in task_description), any query that contains these characters as part of a non-referential word will be misclassified. For example:

• '还款计划查询' ("loan repayment plan query") contains '还' → incorrectly augmented
• '他山之石' (an idiom) contains '他' → incorrectly augmented
• '也门旅游攻略' ("Yemen travel guide") contains '也' → incorrectly augmented

This silently pollutes the embedding input with irrelevant prior-message context, degrading retrieval quality.

Suggestion: For the single-character hints, consider requiring word-boundary context (e.g., check that the character is preceded/followed by whitespace or punctuation), or use jieba tokenization to check at the word level rather than character level.

---

4. src/memos/memories/textual/tree_text_memory/retrieve/task_goal_parser.py (L82-L86)

msg["content"] collected without type-guarding for string, causing potential AttributeError on .strip().

The list comprehension guards with msg.get('content') (a truthiness check), which passes for any non-empty, non-None value — including lists (e.g., multi-modal content blocks like [{"type": "text", "text": "..."}]). The collected value is then assigned to last_user_msg and .strip() is called on it at lines 91/93. If content is a list, .strip() raises AttributeError, crashing the entire retrieval path.

Suggestion: Add an isinstance(msg.get('content'), str) guard in the comprehension filter:

recent_user_msgs = [
    msg["content"]
    for msg in conversation
    if isinstance(msg, dict)
    and msg.get("role") == "user"
    and isinstance(msg.get("content"), str)
    and msg["content"].strip()
]

💡 Suggested Change

Before:

            recent_user_msgs = [
                msg["content"]
                for msg in conversation
                if isinstance(msg, dict) and msg.get("role") == "user" and msg.get("content")
            ]

After:

            recent_user_msgs = [
                msg["content"]
                for msg in conversation
                if isinstance(msg, dict)
                and msg.get("role") == "user"
                and isinstance(msg.get("content"), str)
                and msg["content"].strip()
            ]

---

5. src/memos/memories/textual/tree_text_memory/retrieve/task_goal_parser.py (L98-L102)

_REFERENTIAL_HINTS is a constant reconstructed on every call.

Defining this tuple inside _parse_fast means Python allocates and populates a new tuple object on every invocation. Move it to module level or as a class-level constant to avoid unnecessary allocations and to make it easier to maintain/extend the hint vocabulary.

# At module level
_REFERENTIAL_HINTS = (
    "其他", "别的", "这个", "那个", "它", "他", "她",
    "再", "也", "还", "继续", "上面", "刚说",
    "other", "another", "it", "this", "that", "more",
)

---

6. src/memos/memories/textual/tree_text_memory/retrieve/task_goal_parser.py (L110-L111)

Tail truncation discards the topic anchor of long prior messages.

last_user_msg[-max_ctx:] keeps the last 120 characters. For messages where the topic appears at the beginning (e.g., 'I am looking for hotels near the Eiffel Tower — can you also find me ...'), the subject 'hotels near the Eiffel Tower' is silently dropped, and only the trailing filler text is prepended to the query. This defeats the purpose of the augmentation.

Consider truncating from the front (head truncation) instead, keeping the first 120 characters which typically contain the topic anchor:

last_user_msg = last_user_msg[:max_ctx]

💡 Suggested Change

Before:

                    if len(last_user_msg) > max_ctx:
                        last_user_msg = last_user_msg[-max_ctx:]

After:

                    if len(last_user_msg) > max_ctx:
                        last_user_msg = last_user_msg[:max_ctx]

Generated by cloud-assistant via Open Code Review.

[Memtensor-AI]

❌ Automated Test Results: FAILED

Auto-fix retry 1/2 triggered.

Failed tests:

• test_no_conversation_keeps_original_query
• test_referential_query_gets_augmented
• test_non_referential_long_query_not_augmented
• test_english_referential_query
• test_empty_conversation
• test_conversation_without_user_messages
• test_long_last_user_message_truncated
• test_parse_method_passes_conversation_in_fast_mode

Error details

Tests failed. Failed cases: test_no_conversation_keeps_original_query, test_referential_query_gets_augmented, test_non_referential_long_query_not_augmented, test_english_referential_query, test_empty_conversation

Branch: fix/query-context-drift

[Memtensor-AI]

⚠️ Automated Test Results: INCONCLUSIVE

Automated tests inconclusive (auto-generated test defect); treated as non-blocking. Manual review recommended. Details: The AI-generated test passes a tokenizer keyword argument to TaskGoalParser.__init__(), but the class does not accept such a parameter. [advisory, non-gating] AI-generated tests on branch test/auto-gen-56119b2d819611c1-20260723195034: 62/62 passed — these do NOT affect the PR verdict; review the branch manually.

Branch: fix/query-context-drift