Skip to content

v1.6.0 — ACT-R Activation & Concept-Name Citations

Choose a tag to compare

@mainion-ai mainion-ai released this 09 Apr 10:02
· 295 commits to main since this release

Added

  • ACT-R base-level activation with citation frequency — Wander's base-level activation now follows the ACT-R power-law model: B_i = ln(n) - d·ln(t) where n = citation count + 1, t = age in days, d = 0.5 (standard ACT-R decay). Foundational beliefs cited 28 times receive a ln(28) ≈ 3.3 boost over uncited atoms.

  • Concept-name citation extractor (src/citations.ts) — Discovers informal references between atoms by deriving searchable concept names from atom ID slugs and matching against body text. Three citation layers: explicit relations, atom-ID references, and concept-name references (3.5× larger than atom-ID refs).

    • deriveConceptNames(atomId) — extract searchable keywords from atom slug
    • extractCitations(memoryDir) / indexCitations(memoryDir) — extract and store citations in SQLite
    • mk citations -d <dir> [--json] — CLI command
    • atom_citations SQLite table — Schema bumped to v6.

Changed

  • Sqrt-sigmoid baseBoost — Activation modulation changed from sigmoid to sqrt-sigmoid. Gentler compression preserves activation flow to old hub atoms.
  • relationWeight default: 0.5 → 1.0 — Explicit relation edges now carry ~2× the weight of tag co-occurrence.

Migration

Schema v5 → v6: run mk reindex once after upgrading.

Tests

  • 12 new tests in test/citations.test.ts