@samitouri / QOSami_Agents_Zero-1 / commits / 6ba1f30d

fix: make memory cleanup update stale fragments

Alessandro committed May 11, 2026 at 05:13 UTC 6ba1f30dca2b74622aabfe3e948ba9e5c251ccc8

11 files changed +233 -12

plugins/_memory/helpers/memory.py

+98 -2

@@ -363,11 +363,18 @@ class Memory:

363 )

364

365 async def delete_documents_by_query(

366 - self, query: str, threshold: float, filter: str = ""

366 + self,

367 + query: str,

368 + threshold: float,

369 + filter: str = "",

370 + *,

371 + include_exact: bool = False,

372 + cascade: bool = False,

373 ):

374 k = 100

375 tot = 0

376 removed = []

377 + removed_ids: set[str] = set()

378

379 while True:

380 # Perform similarity search with score

@@ -379,6 +386,7 @@ class Memory:

386 # Extract document IDs and filter based on score

387 # document_ids = [result[0].metadata["id"] for result in docs if result[1] < score_limit]

388 document_ids = [result.metadata["id"] for result in docs]

389 + removed_ids.update(str(doc_id) for doc_id in document_ids)

390

391 # Delete documents with IDs over the threshold score

392 if document_ids:

@@ -392,15 +400,45 @@ class Memory:

400 if len(document_ids) < k:

401 break

402

403 + if include_exact:

404 + exact_docs = self._find_exact_query_docs(query, filter, removed_ids)

405 + if exact_docs:

406 + exact_ids = [doc.metadata["id"] for doc in exact_docs]

407 + await self.db.adelete(ids=exact_ids)

408 + removed += exact_docs

409 + removed_ids.update(str(doc_id) for doc_id in exact_ids)

410 + tot += len(exact_ids)

411 +

412 + if cascade and removed_ids:

413 + related_docs = self._find_related_docs_by_ids(removed_ids)

414 + if related_docs:

415 + related_ids = [doc.metadata["id"] for doc in related_docs]

416 + await self.db.adelete(ids=related_ids)

417 + removed += related_docs

418 + removed_ids.update(str(doc_id) for doc_id in related_ids)

419 + tot += len(related_ids)

420 +

421 if tot:

422 self._save_db() # persist

423 return removed

424

399 - async def delete_documents_by_ids(self, ids: list[str]):

425 + async def delete_documents_by_ids(

426 + self, ids: list[str], *, cascade: bool = False, filter: str = ""

427 + ):

428 # aget_by_ids is not yet implemented in faiss, need to do a workaround

429 rem_docs = await self.db.aget_by_ids(

430 ids

431 ) # existing docs to remove (prevents error)

432 + rem_ids = [doc.metadata["id"] for doc in rem_docs]

433 +

434 + if cascade:

435 + related_docs = self._find_related_docs_by_ids(set(ids) | set(rem_ids))

436 + if related_docs:

437 + existing = {doc.metadata["id"] for doc in rem_docs}

438 + rem_docs.extend(

439 + doc for doc in related_docs if doc.metadata["id"] not in existing

440 + )

441 +

442 if rem_docs:

443 rem_ids = [doc.metadata["id"] for doc in rem_docs] # ids to remove

444 await self.db.adelete(ids=rem_ids)

@@ -445,6 +483,47 @@ class Memory:

483 if not self.db.get_by_ids(doc_id): # check if exists

484 return doc_id

485

486 + def _find_exact_query_docs(

487 + self, query: str, filter: str, skip_ids: set[str]

488 + ) -> list[Document]:

489 + needle = _normalize_memory_match_text(query)

490 + if len(needle) < 3:

491 + return []

492 +

493 + docs: list[Document] = []

494 + comparator = Memory._get_comparator(filter) if filter else None

495 + for doc in self.db.get_all_docs().values():

496 + doc_id = str(doc.metadata.get("id", ""))

497 + if not doc_id or doc_id in skip_ids:

498 + continue

499 + if comparator and not comparator(doc.metadata):

500 + continue

501 + haystack = _normalize_memory_match_text(

502 + f"{doc.page_content}\n{json.dumps(doc.metadata, sort_keys=True, default=str)}"

503 + )

504 + if needle in haystack:

505 + docs.append(doc)

506 + return docs

507 +

508 + def _find_related_docs_by_ids(

509 + self, ids: set[str], filter: str = ""

510 + ) -> list[Document]:

511 + ids = {str(doc_id) for doc_id in ids if str(doc_id)}

512 + if not ids:

513 + return []

514 +

515 + docs: list[Document] = []

516 + comparator = Memory._get_comparator(filter) if filter else None

517 + for doc in self.db.get_all_docs().values():

518 + doc_id = str(doc.metadata.get("id", ""))

519 + if not doc_id or doc_id in ids:

520 + continue

521 + if comparator and not comparator(doc.metadata):

522 + continue

523 + if _metadata_references_any(doc.metadata, ids):

524 + docs.append(doc)

525 + return docs

526 +

527 @staticmethod

528 def _save_db_file(db: MyFaiss, memory_subdir: str):

529 abs_dir = abs_db_dir(memory_subdir)

@@ -547,6 +626,23 @@ def reload():

626 Memory.index = {}

627

628

629 +def _normalize_memory_match_text(value: str) -> str:

630 + return " ".join(str(value or "").casefold().split())

631 +

632 +

633 +def _metadata_references_any(value: Any, ids: set[str]) -> bool:

634 + if isinstance(value, dict):

635 + return any(_metadata_references_any(item, ids) for item in value.values())

636 + if isinstance(value, (list, tuple, set)):

637 + return any(_metadata_references_any(item, ids) for item in value)

638 + text = str(value or "").strip()

639 + if not text:

640 + return False

641 + if text in ids:

642 + return True

643 + return any(doc_id in text.split(",") for doc_id in ids)

644 +

645 +

646 def abs_db_dir(memory_subdir: str) -> str:

647 # patch for projects, this way we don't need to re-work the structure of memory subdirs

648 if memory_subdir.startswith("projects/"):

plugins/_memory/helpers/memory_consolidation.py

+4 -2

@@ -728,9 +728,11 @@ class MemoryConsolidator:

728 updated_count += 1

729 updated_ids.append(new_id)

730

731 - # Step 2: Insert additional new memory if provided

731 + # Step 2: Insert the new memory only when no existing memory was updated.

732 + # UPDATE means "repopulate the existing subject", not "append another

733 + # equally-important memory". This keeps mutable facts from piling up.

734 new_memory_id = None

733 - if result.new_memory_content:

735 + if result.new_memory_content and not updated_ids:

736 # LLM metadata takes precedence over original metadata when there are conflicts

737 final_metadata = {

738 'area': area,

plugins/_memory/prompts/agent.system.memories.md

+2 -1

@@ -1,5 +1,6 @@

1 # Memories on the topic

2 - following are memories about current topic

3 - do not overly rely on them they might not be relevant

4 +- if memories conflict, prefer the newest/current fact and ignore superseded older fragments

5 -{{memories}}

\ No newline at end of file

6 +{{memories}}

plugins/_memory/prompts/agent.system.tool.memory.md

@@ -9,6 +9,10 @@ notes:

9 - `threshold` is similarity from `0` to `1`

10 - `filter` is a metadata expression (e.g. `area=='main'`)

11 - confirm destructive changes when accuracy matters

12 +- when the user updates a durable fact/preference, load related memories first, forget/delete superseded versions, then save one complete current version

13 +- do not append a second memory for the same mutable subject when the new statement replaces the old one

14 +- `memory_forget` also cleans exact matches and derived fragment/solution records related to removed memories

15 +- use `memory_save` for stable current facts, not short-lived test markers, greetings, or one-off conversation events

17 example:

18 ~~~json

plugins/_memory/prompts/memory.consolidation.sys.md

+13 -4

@@ -11,6 +11,7 @@ Analyze a new memory alongside existing similar memories and determine whether t

11 - **keep_separate** if memories serve different purposes

12 - **skip** consolidation if no action is beneficial

14 +Default bias: for mutable user preferences, project state, configuration choices, names, locations, active tasks, or "current" facts about the same subject, prefer **update** or **replace** over appending another separate memory.

16 ## Consolidation Analysis Guidelines

@@ -22,14 +23,16 @@ Analyze a new memory alongside existing similar memories and determine whether t

24 ### 1. Temporal Intelligence

25 - **Newer information** generally supersedes older information

25 -- **Preserve historical context** when consolidating - don't lose important chronological details

26 -- **Consider recency** - more recent memories may be more relevant

26 +- **Preserve historical context** only when the user explicitly needs history or an audit trail

27 +- **Do not keep old preferences as equally important memories** when the new memory clearly gives the current state

28 +- **Consider recency** - more recent memories are usually more relevant for mutable facts

30 ### 2. Content Relationships

31 - **Complementary information** should be merged into comprehensive memories

32 - **Contradictory information** requires careful analysis of which is more accurate/current

33 - **Duplicate content** should be consolidated to eliminate redundancy

34 - **Distinct but related topics** may be better kept separate

35 +- **Same subject, changed value** should usually be update or replace, not keep_separate

37 ### 3. Quality Assessment

38 - **More detailed/complete** information should be preserved

@@ -79,8 +82,8 @@ Provide your analysis as a JSON object with this exact structure:

83 - **merge**: Combine multiple memories into one comprehensive memory, removing originals

84 - **replace**: Replace outdated, incorrect, or superseded memories with new version, preserving important metadata. Use when new information directly contradicts or makes old information obsolete.

82 -- **keep_separate**: New memory addresses different aspects, keep all memories separate

83 -- **update**: Enhance existing memory with additional details from new memory

85 +- **keep_separate**: New memory addresses a genuinely different subject or stable historical event, keep all memories separate

86 +- **update**: Repopulate an existing memory for the same subject with the latest complete current version; do not insert an additional memory when the updated memory is sufficient

87 - **skip**: No consolidation needed, use simple insertion for new memory

89 ## Example Consolidation Scenarios

@@ -95,6 +98,11 @@ Provide your analysis as a JSON object with this exact structure:

98 **Existing**: "User API endpoint is /api/users for getting user data"

99 **Action**: replace → Update with new endpoint, note the change in historical_notes

100

101 +### Scenario 2b: Update Current Preference

102 +**New**: "User now prefers concise technical answers with examples"

103 +**Existing**: "User prefers long exploratory answers"

104 +**Action**: update -> Rewrite the existing user-preference memory to the new current preference. Do not keep both as equally relevant memories.

105 +

106 **REPLACE Criteria**: Use replace when:

107 - **High similarity score** (>0.9) indicates very similar content

108 - New information directly contradicts existing information

@@ -116,6 +124,7 @@ Provide your analysis as a JSON object with this exact structure:

124 3. **Maintain Context**: Keep temporal and source information where relevant

125 4. **Enhance Searchability**: Use consolidation to improve future memory retrieval

126 5. **Reduce Redundancy**: Eliminate unnecessary duplication while preserving nuance

127 +6. **Keep Current Facts Current**: For mutable facts, the final memory should represent the latest usable state, not a human-like archive of every old version

128

129 ## Instructions

130

plugins/_memory/prompts/memory.memories_filter.sys.md

+4 -1

@@ -13,6 +13,8 @@

13 - Focus on USER MESSAGE if provided, use HISTORY for context

14 - Keep in mind that these memories should be helpful for continuing the conversation and solving problems by AI

15 - Consider if each memory holds real information value for the context or not

16 +- If multiple memories conflict about the same mutable user/project fact, include only the newest/current one when it is identifiable

17 +- Exclude superseded, historical, duplicate, or low-detail fragments when a more complete current memory is available

19 # Include only when:

20 - Memory is relevant to the current situation

@@ -22,6 +24,7 @@

24 - Short vague texts like "Pet inquiry" or "Programming skills" with no more detail

25 - Common conversation patterns like greetings

26 - Memories that hold no information value

27 +- Older conflicting memories for the same preference or project state when a newer/current memory is available

29 # Example output

30 ```json

@@ -32,4 +35,4 @@

35 > "User has greeted me" (no information value)

36 > "Hello world program" (just title, no details, no context, irrelevant by itself)

37 > "Today is Monday" (just date, information obsolete, not helpful)

35 -> "Memory search" (just title, irrelevant by itself)

\ No newline at end of file

38 +> "Memory search" (just title, irrelevant by itself)

plugins/_memory/prompts/memory.memories_sum.sys.md

@@ -30,6 +30,9 @@

30 - Do not break information related to the same subject into multiple memories, keep them as one text

31 - If there are multiple facts related to the same subject, merge them into one more detailed memory instead

32 - Example: Instead of three memories "User's dog is Max", "Max is 6 years old", "Max is white and brown", create one memory "User's dog is Max, 6 years old, white and brown."

33 +- If the history changes or corrects a previously stated fact, output only the new complete current fact; do not output both old and new versions

34 +- Prefer a single durable profile-style sentence for mutable user/project preferences, such as "User currently prefers..." or "Project currently uses..."

35 +- Do not memorize temporary test markers, temporary behavior checks, or cleanup-only facts

37 # Correct examples of data worth memorizing with (explanation)

38 > User's name is John Doe (name is important)

@@ -45,6 +48,8 @@

48 > Today is Monday (just date, no value in this information)

49 > Market inquiry (just a topic without detail)

50 > RAM Status (just a topic without detail)

51 +> User used to prefer X before changing to Y (historical preference is usually not useful; memorize the current preference only)

52 +> Temporary marker ABC123 was used in a memory test (test residue, not useful)

55 # Further WRONG examples

plugins/_memory/tools/memory_delete.py

+1 -1

@@ -8,7 +8,7 @@ class MemoryDelete(Tool):

8 async def execute(self, ids="", **kwargs):

9 db = await Memory.get(self.agent)

10 ids = [id.strip() for id in ids.split(",") if id.strip()]

11 - dels = await db.delete_documents_by_ids(ids=ids)

11 + dels = await db.delete_documents_by_ids(ids=ids, cascade=True)

13 result = self.agent.read_prompt("fw.memories_deleted.md", memory_count=len(dels))

14 return Response(message=result, break_loop=False)

plugins/_memory/tools/memory_forget.py

+7 -1

@@ -8,7 +8,13 @@ class MemoryForget(Tool):

9 async def execute(self, query="", threshold=DEFAULT_THRESHOLD, filter="", **kwargs):

10 db = await Memory.get(self.agent)

11 - dels = await db.delete_documents_by_query(query=query, threshold=threshold, filter=filter)

11 + dels = await db.delete_documents_by_query(

12 + query=query,

13 + threshold=threshold,

14 + filter=filter,

15 + include_exact=True,

16 + cascade=True,

17 + )

19 result = self.agent.read_prompt("fw.memories_deleted.md", memory_count=len(dels))

20 return Response(message=result, break_loop=False)

tests/test_memory_cleanup.py new

+93

@@ -0,0 +1,93 @@

1 +from __future__ import annotations

2 +

3 +import sys

4 +import asyncio

5 +from pathlib import Path

6 +

7 +from langchain_core.documents import Document

8 +

9 +PROJECT_ROOT = Path(__file__).resolve().parents[1]

10 +if str(PROJECT_ROOT) not in sys.path:

11 + sys.path.insert(0, str(PROJECT_ROOT))

12 +

13 +from plugins._memory.helpers.memory import Memory

14 +

15 +

16 +class FakeFaiss:

17 + def __init__(self, docs: list[Document]):

18 + self.docs = {doc.metadata["id"]: doc for doc in docs}

19 + self.deleted: list[str] = []

20 +

21 + async def asearch(self, *_args, **_kwargs):

22 + return []

23 +

24 + async def adelete(self, ids):

25 + for doc_id in ids:

26 + self.deleted.append(doc_id)

27 + self.docs.pop(doc_id, None)

28 +

29 + async def aget_by_ids(self, ids):

30 + return [self.docs[doc_id] for doc_id in ids if doc_id in self.docs]

31 +

32 + def get_all_docs(self):

33 + return self.docs

34 +

35 + def get_by_ids(self, ids):

36 + return [self.docs[doc_id] for doc_id in ids if doc_id in self.docs]

37 +

38 +

39 +def test_memory_forget_removes_exact_matches_and_derived_fragments():

40 + main = Document(

41 + page_content="User currently prefers memory cleanup token banana-397.",

42 + metadata={"id": "main-1", "area": "main"},

43 + )

44 + fragment = Document(

45 + page_content="Derived note from old preference.",

46 + metadata={

47 + "id": "fragment-1",

48 + "area": "fragments",

49 + "consolidated_from": ["main-1"],

50 + },

51 + )

52 + unrelated = Document(

53 + page_content="Unrelated memory about project setup.",

54 + metadata={"id": "other-1", "area": "main"},

55 + )

56 + fake_db = FakeFaiss([main, fragment, unrelated])

57 + memory = Memory(fake_db, memory_subdir="test")

58 + memory._save_db = lambda: None

59 +

60 + removed = asyncio.run(

61 + memory.delete_documents_by_query(

62 + query="banana-397",

63 + threshold=0.99,

64 + include_exact=True,

65 + cascade=True,

66 + )

67 + )

68 +

69 + assert {doc.metadata["id"] for doc in removed} == {"main-1", "fragment-1"}

70 + assert fake_db.deleted == ["main-1", "fragment-1"]

71 + assert set(fake_db.docs) == {"other-1"}

72 +

73 +

74 +def test_memory_delete_cascades_even_when_original_id_is_already_missing():

75 + replacement = Document(

76 + page_content="User currently prefers concise technical answers.",

77 + metadata={

78 + "id": "replacement-1",

79 + "area": "main",

80 + "updated_from": "old-pref-1",

81 + },

82 + )

83 + fake_db = FakeFaiss([replacement])

84 + memory = Memory(fake_db, memory_subdir="test")

85 + memory._save_db = lambda: None

86 +

87 + removed = asyncio.run(

88 + memory.delete_documents_by_ids(["old-pref-1"], cascade=True)

89 + )

90 +

91 + assert [doc.metadata["id"] for doc in removed] == ["replacement-1"]

92 + assert fake_db.deleted == ["replacement-1"]

93 + assert fake_db.docs == {}

tests/test_tool_action_contracts.py

@@ -173,6 +173,8 @@ def test_memory_forget_tool_imports_plugin_memory_load(monkeypatch):

173 "query": "codex memory forget token",

174 "threshold": 0.99,

175 "filter": "area=='codex_sweep'",

176 + "include_exact": True,

177 + "cascade": True,

178 }

179 ]

180