Context: memory-lancedb-pro 1.1.0-beta.11, deployed per SKILL.md Plan A config.
Summary
The Plan A config recommends minScore: 0.6, hardMinScore: 0.62 with cross-encoder rerank enabled. On current plugin releases this combination structurally excludes BM25-only results: results found only by full-text search enter hybrid fusion with a floor score of 0.5, which can never clear a hardMinScore of 0.62. Short natural queries against stored entries also commonly score below 0.62 on the vector side.
Reproduction on a live store (203 entries)
openclaw memory-pro search "desk plant" --scope agent:X returned "No relevant memories found" while the store contained an entry whose text literally begins "User desk plant care: snake plant named ...". The same entry is returned at 79% for a denser query. After changing thresholds to 0.45 / 0.35 the same short query returns the entry as the top hit at 59%, which is exactly the band the 0.62 bar was rejecting.
Three official sources currently disagree
| Source |
minScore |
hardMinScore |
| SKILL.md Plan A (rerank enabled) |
0.6 |
0.62 |
| SKILL.md Plan B (no rerank) |
0.3 |
0.35 |
| setup-memory.sh installer (pro-rerank level) |
0.45 |
0.35 |
The installer's scaling looks correct: with a cross-encoder as the precision layer, pre-filter thresholds should be lower, not higher. Plan A inverts that relationship and silently disables the lexical half of hybrid retrieval.
Suggestion
Align Plan A with the installer's rerank profile (0.45 / 0.35), or add a warning that hardMinScore must stay below the BM25-only floor (0.5 on current releases; note the main repo's unreleased changelog removes the floor, which makes high hardMinScore values even more restrictive for lexical-only hits, not less).
Context: memory-lancedb-pro 1.1.0-beta.11, deployed per SKILL.md Plan A config.
Summary
The Plan A config recommends
minScore: 0.6, hardMinScore: 0.62with cross-encoder rerank enabled. On current plugin releases this combination structurally excludes BM25-only results: results found only by full-text search enter hybrid fusion with a floor score of 0.5, which can never clear a hardMinScore of 0.62. Short natural queries against stored entries also commonly score below 0.62 on the vector side.Reproduction on a live store (203 entries)
openclaw memory-pro search "desk plant" --scope agent:Xreturned "No relevant memories found" while the store contained an entry whose text literally begins "User desk plant care: snake plant named ...". The same entry is returned at 79% for a denser query. After changing thresholds to 0.45 / 0.35 the same short query returns the entry as the top hit at 59%, which is exactly the band the 0.62 bar was rejecting.Three official sources currently disagree
The installer's scaling looks correct: with a cross-encoder as the precision layer, pre-filter thresholds should be lower, not higher. Plan A inverts that relationship and silently disables the lexical half of hybrid retrieval.
Suggestion
Align Plan A with the installer's rerank profile (0.45 / 0.35), or add a warning that hardMinScore must stay below the BM25-only floor (0.5 on current releases; note the main repo's unreleased changelog removes the floor, which makes high hardMinScore values even more restrictive for lexical-only hits, not less).