Found while auditing overview.md in PR #239 (fixed there, but the same drift exists on other pages, out of scope for a single-page PR).
Issue 1 — stale "9-component" count
core-concepts/hybrid-search.md was corrected to "10-Component Scoring System" in PR #194/#202, and architecture/data-stores.md already says "10-component". But these pages still say 9-component:
architecture/overview.md — "### 9-Component Hybrid Scoring"getting-started/introduction.md — "...re-ranks results using a 9-component relevance score"core-concepts/relationship-types.md — Score["Hybrid Scoring<br/>9-component"]reference/api/recall-operations.md — "combines nine scoring components" / "9-component scoring system" / "### 9-Component Score Calculation"
Actual formula (automem/utils/scoring.py, confirmed at automem@4b5eaafd2602c9eba39bbfe38e4120e3654c67e9) sums exactly 10 weighted components: vector, keyword, metadata, tag, importance, confidence, recency, exact, relation, plus a context/relevance bonus (SEARCH_WEIGHT_RELEVANCE).
Issue 2 — undocumented SEARCH_WEIGHT_TEMPORAL
automem/config.py also defines an 11th weight, SEARCH_WEIGHT_TEMPORAL (default 0.1), gating a relative-recency re-rank pass (RECALL_RECENCY_BIAS env or a per-request recency_bias param — see issues #158/#159 in the automem repo). It's inert unless that re-rank runs, so it's separate from the base 10-component formula, but it isn't mentioned in any doc reviewed so far (core-concepts/hybrid-search.md's "Configuration Reference" weight table doesn't list it either).
Ask: a maintainer pass to (a) bump the 4 "9-component" pages to 10, and (b) decide whether/where to document SEARCH_WEIGHT_TEMPORAL / recency_bias — is this a real user-facing knob worth its own section, or still experimental?
Verified against: automem@4b5eaafd2602c9eba39bbfe38e4120e3654c67e9
Found while auditing
overview.mdin PR #239 (fixed there, but the same drift exists on other pages, out of scope for a single-page PR).Issue 1 — stale "9-component" count
core-concepts/hybrid-search.mdwas corrected to "10-Component Scoring System" in PR #194/#202, andarchitecture/data-stores.mdalready says "10-component". But these pages still say 9-component:architecture/overview.md— "### 9-Component Hybrid Scoring"getting-started/introduction.md— "...re-ranks results using a 9-component relevance score"core-concepts/relationship-types.md—Score["Hybrid Scoring<br/>9-component"]reference/api/recall-operations.md— "combines nine scoring components" / "9-component scoring system" / "### 9-Component Score Calculation"Actual formula (
automem/utils/scoring.py, confirmed atautomem@4b5eaafd2602c9eba39bbfe38e4120e3654c67e9) sums exactly 10 weighted components: vector, keyword, metadata, tag, importance, confidence, recency, exact, relation, plus a context/relevance bonus (SEARCH_WEIGHT_RELEVANCE).Issue 2 — undocumented
SEARCH_WEIGHT_TEMPORALautomem/config.pyalso defines an 11th weight,SEARCH_WEIGHT_TEMPORAL(default0.1), gating a relative-recency re-rank pass (RECALL_RECENCY_BIASenv or a per-requestrecency_biasparam — see issues #158/#159 in theautomemrepo). It's inert unless that re-rank runs, so it's separate from the base 10-component formula, but it isn't mentioned in any doc reviewed so far (core-concepts/hybrid-search.md's "Configuration Reference" weight table doesn't list it either).Ask: a maintainer pass to (a) bump the 4 "9-component" pages to 10, and (b) decide whether/where to document
SEARCH_WEIGHT_TEMPORAL/recency_bias— is this a real user-facing knob worth its own section, or still experimental?Verified against: automem@4b5eaafd2602c9eba39bbfe38e4120e3654c67e9