feat: prevent repeated memories via archival and activated-only retrieval

[CaralHsi]

Summary

This PR closes the loop on the “repeated memory” problem at the add stage.

Previously, when adding new memories, the system could:

• create a new node that is semantically identical to an existing one,
• continue to return both the old and the new memory during recall,
• allow merged_from memories to keep influencing downstream reasoning.

Now, when a new memory is added, we:

1. search for similar existing memories,
2. let a dedicated “memory merge” prompt decide whether they should be merged,
3. archive any merged_from nodes so they no longer participate in recall.

Combined with stricter filtering on status="activated", this ensures that:

• stable facts are stored once,
• merged-away memories are cleanly retired,
• recall results are less noisy and more consistent.

---

What’s changed

1. Mem-reader: detect & merge repeated memories on add

• Inject graph_db into MemReader via MemReaderFactory.from_config(..., graph_db=...), so readers can:
  • look up similar memories before writing,
  • perform semantic deduplication and conflict detection.
• Add dedicated memory merge prompts (MEMORY_MERGE_PROMPT_EN / MEMORY_MERGE_PROMPT_ZH) to:
  • split the new input into fact units (identity, stable preferences, relationships, one-off plans),
  • distinguish “repeated confirmation of the same fact” from genuinely new information,
  • decide whether to:
    • merge and return { value, merged_from, should_merge: true }, or
    • skip merging with { should_merge: false }.
• When merging, the new memory:
  • keeps a single, non-redundant value,
  • records the source IDs in merged_from for later archival.

2. Graph DB & recall: only activated memories, plus archival

• Extend the graph DB base and implementations (Neo4j / PolarDB) to support:
  • update_node(id, fields, user_name=None) with tenant context,
  • get_node(..., **kwargs) and get_nodes(ids, ..., **kwargs) for user-scoped queries,
  • get_by_metadata(filters, status=None) and get_all_memory_items(scope, ..., status=None) with optional status filtering.
• In TreeText retrieval:
  • all metadata-based lookups now filter with status="activated",
  • all bulk node loads pass user_name to respect multi-tenant isolation.
• In scheduler & single-cube flows:
  • whenever a memory carries merged_from in metadata.info, the referenced IDs are updated to:

    {"status": "archived"}

    with the proper user_name / mem_cube_id,
  • archived nodes are thus excluded from future recall.

3. Prompt & behavior: make “repeat vs. new” explicit

• Introduce explicit, language-level rules for memory merging:
  • do not repeat the same fact multiple times or just because it was mentioned at different timestamps,
  • avoid mixing long-term preferences with one-off plans in the same merged value,
  • only include time information when it changes the meaning (e.g. “used to dislike → now likes”).
• This makes the add-path behavior much more predictable:
  • stable user facts converge into a single clean memory,
  • purely one-off events are not mistakenly merged into long-term profiles,
  • downstream components see a de-duplicated, clearer memory graph.

Fix: #881

Docs Issue/PR: (docs-issue-or-pr-link)

Reviewer: @(reviewer)

Checklist:

• I have performed a self-review of my own code | 我已自行检查了自己的代码
• I have commented my code in hard-to-understand areas | 我已在难以理解的地方对代码进行了注释
• I have added tests that prove my fix is effective or that my feature works | 我已添加测试以证明我的修复有效或功能正常
• I have created related documentation issue/PR in MemOS-Docs (if applicable) | 我已在 MemOS-Docs 中创建了相关的文档 issue/PR（如果适用）
• I have linked the issue to this PR (if applicable) | 我已将 issue 链接到此 PR（如果适用）
• I have mentioned the person who will review this PR | 我已提及将审查此 PR 的人