Mémoires

The Arcueil catalog of Bayesian craft — curated, evidence-graded, contradiction-aware. Named for the Mémoires de la Société d'Arcueil*, and read by jaynes-robot. License: CC BY 4.0.*

Why this exists

When your hierarchical model diverges, you search the Stan forum with three keywords and read five half-relevant threads. Search this catalog instead: the same community knowledge — ≈27.5k forum threads (Stan/PyMC/Pyro) + the Betancourt and Simpson corpora + a human-curated pymc-labs peer layer — distilled, adversarially reviewed, evidence-graded, and organized so that every answer comes with cross-cutting evidence: the same challenge in other model classes, the same model in other libraries, and the principle that explains why.

Use it

pip install memoires        # zero-dependency core (stdlib sqlite FTS5)

memoires search "divergences hierarchical funnel"   # the search engine
memoires show  hierarchical-multilevel/C2            # one entry, with all its edges
memoires graph hierarchical-multilevel/P4            # ↑ claims · ↔ related · sources
memoires stats

Semantic search (finds concepts, not just keywords — "sampler slow but no divergences" → the geometry-vs-sampler principle):

pip install 'memoires[semantic]'          # adds a frozen embedder; the 768-d index ships in the wheel
memoires search "my chain won't converge" # hybrid lexical+semantic by default when installed
memoires search --semantic "funnel neck geometry"

Also browsable entirely on GitHub — every entry is a markdown node with clickable edges.

Repository structure

memoires/    the pip package: the catalog (claims/recs/super-axioms + graph data) and the CLI —
             a local search engine, and the foundation for the MCP server agents will use
process/     how it is built: pipeline, audits, methodology, experiments, provenance,
             worklog (the forking path), roadmap, update procedure
data/        raw sources (gitignored) + one provenance note per source (tracked)

The evidence graph

Nothing here asks to be trusted on authority — every entry is a node in an explicit, clickable graph, so you can always see where it comes from:

7  super-axioms      the leanest why                 SUPER_AXIOMS.md
      ↑ subsumes (82/82, verified bijection)
82 claims            mid-level principles            memoires/catalog/claims/<page>/<C-id>.md
      ↑ grounds (613 edges; 27 honest gaps)
640 recs             "for model X / when you see Y   memoires/catalog/recs/<page>/<id>.md
                      → works ✓ / doesn't ✗ (+conditions)"
      → sources      307 short-ids, ALL resolving    memoires/catalog/data/source_map.json
                     to clickable forum threads,
                     blog case studies, pymc-labs skills
      ↔ related      2,886 cross-page links          "similar challenge, other models"

Every claim file: statement → nuance → conditions → tier, its super-axiom (↑), the recs it grounds (↓), clickable sources, and related entries across the catalog. Every rec file: the ✓/✗ verdict with its conditions, the governing claim(s) (↑), attached diagnostic moves, an efficacy slot for empirical grounding, clickable sources, and cross-cutting neighbors.

How to read it

• Top-down — SUPER_AXIOMS.md → follow subsumes-links down into claims → practical evidence.
• By your model — the pages below are index/summary views; every entry links to its full node.
• By your symptom — the cross-cutting pages are indexed by what you actually see (divergences, R̂ alarms, treedepth, prior doubt).
• The spine — CLAIMS_SPINE.md, all 82 claims on one page.

Cross-cutting (computation & diagnostics — apply across all models)

• Posterior geometry & gradient-based sampling
• Priors & identifiability
• Model evaluation
• Convergence & Monte-Carlo reliability diagnostics
• Decision theory (seeded — thin by honest design)

By model class (a navigation tag, not the only axis)

Regression · Hierarchical / multilevel · Mixtures · Gaussian processes · Time series & state space · Spatial & areal · Latent factor · ODE / dynamical · Measurement error & missingness · Sparse regression & shrinkage

Trust, stated plainly

The catalog's own quality is evidence-graded, like its contents: claims reviewed for over-generalization (75/76 faithful, the 1 fixed), the error-dense rec class rigorously swept, a 50-entry adversarial release review with double independent verification (10 defects found and fixed), and an external cross-check against human-curated expert material in which 5 of 7 disagreements resolved in this catalog's favor. Full methodology and bounds: PROVENANCE.md · audits/ · methodology/. Known limitations: GAPS.md.

« La théorie des probabilités n'est, au fond, que le bon sens réduit au calcul. » — Laplace