WIP: drop importance_score and confidence columns from memories#55
Merged
Conversation
Schema migration only — no application-code changes yet. Application side will follow in subsequent commits. Both columns carried mechanical defaults (importance_score=1.0, confidence=0.5 for observations) and were no-ops in recall scoring. Industry consensus (mem0, Letta, LangMem, Zep, Cognee) is to skip per-memory quality scores entirely. Lore's existing cosine + FTS hybrid recall handles ranking without them. Indexes are dropped before columns to satisfy SQLite's DROP COLUMN restriction. The PG on-access trigger (if present) is also dropped. Out of scope (intentionally untouched): per-relationship confidence on graph entities/mentions/relationships.
Removed from: * NewMemory.confidence (default 0.5) * NewMemory.importance_score (Optional[float]) * StoredMemory.confidence * StoredMemory.importance_score * MemoryPatch.confidence * ExportedMemory.confidence Persistence implementations (sqlite/postgres) and all callers will follow in subsequent commits — until they're updated, those modules will fail to import (unused-attribute access on the dataclasses). NewMention.confidence and StoredMention.confidence (graph relationship quality) are intentionally untouched — separate column on a separate table.
* INSERT statements no longer write the columns * SELECT lists drop them everywhere they appeared * _row_to_memory and _row_to_exported stop populating the dropped fields * recall_by_embedding score formula drops the importance multiplier (was a no-op since every row had identical 1.0) * bump_access_counts and record_memory_access UPDATE drops the importance_score recomputation * list_candidate_memories_for_recommendation now orders by created_at DESC (importance_score column is gone) * search_memories_text orders by created_at DESC * GraphStats loses the avg_importance field — also removed from the dataclass in types.py * AVG(importance_score) aggregate queries removed from get_graph_stats * _MEMORY_COLS constant trimmed * ScoredMemory mapping no longer copies the dropped fields Graph-table per-mention confidence (entity_mentions.confidence) is unaffected — different table, different concept.
…tence * postgres.py: column lists, RETURNING clauses, recall scoring, bump_access_counts UPDATE, AVG aggregates, search/recall/timeline queries all stop carrying the dropped columns. Recall score formula drops the importance multiplier (was a no-op). * protocol.py: import_extracted_memory and upsert_memory_with_embedding signatures lose the confidence parameter; bump_access_counts docstring updated. * sqlite.py: matching shape changes for upsert_memory_with_embedding and import_extracted_memory. * GraphStats.avg_importance removed in types.py; PG and SQLite graph-stats methods stop computing it. * list_candidate_memories_for_recommendation now orders by created_at DESC (importance ordering becomes meaningless once the column is gone). The lone remaining `m.confidence` reference in postgres.py (entity_mentions INSERT) is the per-relationship graph confidence on a separate table — intentional, out of scope.
…re.py * Delete src/lore/importance.py (compute_importance, time_adjusted_importance no longer needed — importance_score column gone) * New src/lore/decay.py with decay_factor + resolve_half_life (the two pure functions still needed for recall scoring) * lore.py: replace time_adjusted_importance with new _memory_decay helper (recency decay only, no importance multiplier) * lore.py: upvote/downvote stop recomputing importance_score * lore.py: cleanup_expired drops the importance-threshold filter and uses decay-only thresholding * lore.py: recalculate_importance method removed (no column to recompute) * Delete tests/test_importance_scoring.py (whole file was about the dropped functions)
…sons services * observations.py: drop the hardcoded confidence=0.5 / importance_score=0.5 from create_observation * memories.py: drop confidence parameter from create_memory and update_memory * lessons.py: drop confidence param from create() and update(); rename min_confidence to min_score in search() (it was always a min-final-score threshold, despite the name); strip importance_score / confidence from search response, record_access response, and bulk-upsert path Service-layer callers (HTTP routes, MCP, CLI, tests) will be updated in subsequent commits; until they catch up they'll fail with TypeError on the dropped kwargs.
Records what's already done (Phases 1-5a, 6 commits) and what's left (Phases 5b-10) so a fresh session can resume cleanly. Includes the "out of scope" boundaries (graph mention confidence, recommendation confidence, freshness confidence, classification confidence) so the executor doesn't over-reach.
amitpaz1
force-pushed
the
fix/drop-quality-score-columns
branch
from
54859be to
53c408f
Compare
amitpaz1
changed the title
- services/conversations.py: drop confidence pass-through to import_extracted_memory
- services/snapshots.py: drop confidence=1.0 from NewMemory
- services/retrieve.py: drop importance multiplier (always 1.0 in data) and importance signal
- services/graph/review.py: drop source_importance from risk score; RiskScore loses source_reliability
- services/graph/graph.py: drop importance/confidence from GraphNode and SearchHit; drop min_importance from get_graph_data
- server/routes/graph/{models,memories,stats}.py: drop matching response fields and query params
- consolidation.py: replace importance-based fallback with most-recent (created_at); drop fields from new Memory
- conversation/extractor.py: stop forwarding candidate confidence to lore.remember
- ingest/dedup.py + store/http.py: rename min_confidence -> min_score for the search() RPC
LEAVE entity_mention/relationship confidence and Fact.confidence (graph concepts)
LEAVE recommend / freshness / classify confidence (separate concepts)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…odels, and routes MCP (mcp/server.py): - drop importance: ... from recall and list_memories text formatters - update on_this_day docstring (was "sorted by importance") HTTP layer (server/models.py and routes): - LessonCreate/Response/Export/Import: drop confidence field - LessonSearchRequest, MemorySearchRequest: rename min_confidence -> min_score - MemoryCreate/Response/Update: drop confidence field - routes/lessons.py: drop confidence/min_confidence/importance_score plumbing - routes/memories.py: drop confidence param, drop importance_score from access response, fix update validator - routes/temporal.py: drop confidence from MemoryResponse - routes/recent.py: drop importance_score field and detailed-format suffix - routes/retention.py: drop min_importance_score query param + AffectedMemory.importance_score - routes/review.py: drop source_reliability from RiskScore (matches services side) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… export, UI
Top-level:
- lore.py: drop confidence param from remember(); rename min_confidence -> min_score
in recall (delegates) and _recall_local (now filters on score, not memory column);
rename importance_threshold -> decay_threshold (the column it referenced is gone);
cleanup_expired now decays-only.
- async_lore.py: drop confidence from remember(); drop recalculate_importance() (no-op
after the column was removed); cleanup_expired keyword renamed for parity; drop
confidence from PromptFormatter Memory adapter.
- types.py: drop importance_score, confidence from Memory dataclass; drop
avg_importance / below_threshold_count from MemoryStats.
- temporal.py: drop importance_score from on_this_day sort key and format output.
- recent.py: drop importance_score from detailed/structured formatters.
- retention.py: drop min_importance_score from RetentionPolicy; _find_expired now
age-only; archive output drops importance_score; module docstring updated.
- store/http.py: drop confidence from create payload, response parser, update path.
CLI:
- cli/__init__.py: drop --confidence and --min-importance CLI args; drop --sort importance.
- cli/commands/remember.py, manage.py, misc.py: drop matching arg plumbing and the
importance column from list/json output.
Export:
- export/serializers.py: drop importance_score / confidence from memory_to_dict + dict_to_memory.
- export/markdown.py: drop the same fields from per-memory frontmatter.
UI source (note: dist/app.js is a build artifact and needs re-build):
- ui/src/panels/detail.js: drop the Importance and Confidence rows.
- ui/src/panels/stats.js: drop avg_importance row.
- ui/src/panels/filters.js: drop the min-importance slider, unused debounce import.
- ui/src/state.js: drop minImportance from filters / URL state / matcher.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… regression test, repair lessons view
Migration fix:
- migrations/025_drop_quality_score_columns.sql: DROP VIEW lessons CASCADE
before the ALTER TABLE so the rewrite-rule-bearing view is rebuilt; recreate
the view (and INSERT/UPDATE/DELETE rules) without the dropped columns so any
legacy /v1/lessons callers keep working.
- migrations_sqlite/025_drop_quality_score_columns.sql: same fix for the
SQLite read-only view; was failing migration on a fresh DB
("error in view lessons after drop column: no such column: importance_score").
Tests:
- Drop importance_score / confidence args from StoredMemory / NewMemory /
ScoredMemory / ExportedMemory / Memory factories across ~30 test files.
- Drop tests that exercised behaviours we removed: confidence validation,
recalculate_importance noop, source_reliability risk component,
importance-only ordering of recommendation candidates, vote-factor recall
boost (no longer in the multiplicative formula).
- Update test_consolidation full pipeline to use distinct created_at so the
most-recent-wins fallback is deterministic; bump both timestamps past the
short-tier 7-day eligibility threshold.
- Update test_temporal to assert year-DESC + created_at-DESC instead of
importance ordering.
- Rename min_confidence kwarg to min_score in HTTP / lessons-search tests.
- Add tests/persistence/test_quality_columns_dropped.py regression test:
asserts SqliteStore.open() yields a memories table without
importance_score / confidence and that ``from lore import importance``
raises ImportError.
Code:
- Drop the now-unused `Optional` import flagged by ruff.
verification:
- PYTHONPATH=src pytest tests/ --ignore=tests/integration → 2706 passed,
1177 skipped.
- ruff check src/ tests/ → clean.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…y + assertions
CI exposed two missed sites in tests/integration/test_remote.py (the local
suite skips this directory):
- _stored_lesson_row helper: drop the legacy ``"confidence": 0.8`` row field;
the wire shape no longer includes it.
- test_full_flow_publish_query_verify: drop the ``data["confidence"] == 0.8``
assertion against the GET /v1/lessons/{id} response.
- test_export_import_between_contexts: drop the ``confidence=0.8`` kwargs from
both ExportedMemory constructors and the corresponding round-trip POST body.
Verified: PYTHONPATH=src pytest tests/ → 2787 passed, 1205 skipped.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Status: WIP, ~50% done. Mid-refactor checkpoint. Ready for a fresh session to resume.
The full plan + remaining-work checklist is committed at:
docs/superpowers/plans/2026-05-09-drop-quality-score-columns.mdWhat's done (8 commits)
migrations/025_drop_quality_score_columns.sql+ SQLite version) — DROP TRIGGER, DROP INDEX, DROP COLUMN. Verified parses and drops both columns + indexes correctly.persistence/types.py)src/lore/importance.pydeleted; replaced bysrc/lore/decay.pywith the still-needed pure functions (decay_factor,resolve_half_life);lore.pycleansed (time_adjusted_importance→_memory_decay);tests/test_importance_scoring.pydeletedconfidenceparameter, dropimportance_scorefrom response dicts, renamemin_confidence→min_scoreWhat's left
remember())Don't over-reach: leave these "confidence" concepts untouched
confidence/weight(entities/mentions/relationships tables)recommend/engine.pyRecommendationConfidencefreshness/detector.pystaleness confidenceclassify/axis confidencegrep -rn "importance_score\|memory\.confidence" src/ | grep -v "graph\|recommend\|freshness\|classify"should drive the cleanup; don't touch refs in those four directories.Tests will be RED until Phase 5b–9 land
Until services / routes / tests are updated, the suite will fail with TypeError on dropped kwargs. That's expected at this checkpoint — DO NOT panic and revert.
🤖 Generated with Claude Code