limbic 0.1.0

limbic v0.1.0

First feature-complete release. Not published to npm — the limbic name is
unclaimed; install from a checkout (see the README's Install section).

What it is

A local-first, emotion-aware, LLM-agnostic memory engine for TypeScript:
importance-weighted scoring, per-category forgetting curves, GIST
diversity-aware retrieval (arXiv:2405.18754) and LLM extraction — zero required
runtime dependencies, Node ≥ 20, dual ESM/CJS.

Highlights

• createLimbic({ store, embedder, complete, weights, lambda, pool })
  returning remember / extract / retrieve / decayPass / store.
• Stores: zero-dependency MemStore default; SqliteStore behind the
  optional better-sqlite3 peer, held to the same contract suite.
• Embedders: OllamaEmbedder (POST /api/embed), plus node-llama-cpp
  and transformers.js adapters loaded through dynamic import() with an
  install hint. No embedder means keyword-only scoring — the cosine channel is
  MISSING, never 0.
• Diversity conformance: divsel's golden-selection.json — 22 of 22
  cases pass, none skipped, with exact numeric agreement (worst
  tolerance-budget share 0), because the distance kernel reproduces divsel's
  f32 16-accumulator arithmetic via Math.fround.
• Scoring parity: the origin engine's golden-scoring.json — base score,
  final score and both ranking arrays at the fixture's own absolute 1e-6,
  clock pinned to the fixture's now.
• Both fixtures are sha256-pinned and re-hashed on every test run; provenance
  in test/fixtures/PROVENANCE.md.

Notes on the port

• lambda defaults to 0.5 (the origin engine's default); divsel's own
  default is 1.0. The bundled bench shows the crossover between them on a
  clustered corpus — limbic's default is the conservative one.
• The origin engine's LLM extraction path returns [] today due to a prompt
  formatting bug it swallows; limbic substitutes the placeholder literally
  instead.
• Unknown extraction types are kept and mapped to general rather than
  dropped.

Full details in
CHANGELOG.md.