PyAutoMemory is the Memory of the PyAutoScientist — the long-term memory
of the organism: what it has learned, distilled into cross-linked LLM wikis —
literature summaries, scientific concepts, and the citation metadata to
verify them. Memory holds what the science says; what the organism did
lives in the Mind. Source PDFs live off-repo; what's here is the durable
knowledge. Start at index.md.
See the PyAutoMemory Dashboard
for managing all of it from one page: the reading
queue, the paper sections still needing a canonical citation key, and each
sub-wiki's maturity — every work queue carrying a one-tap 📋 button that
copies a paste-ready AI assistant prompt (file the next paper, resolve keys,
upgrade a stub, or Use the memory skill. <domain> to recall what's known). Every queued
paper links to its arXiv abstract page and carries a 📄 button onto the PDF
itself, so a phone can collect a stack of papers to read offline in one tap
each.
One-tap mode. Out of the box each action button opens a prefilled GitHub issue you then submit. Tap the 🔑 chip under the header once and paste a fine-grained token (Resource owner PyAutoLabs, repository PyAutoMemory only, Issues: read and write): from then on every button files that same issue from the page, the row disappears at once, and 📥/📑 ask for your notes in an inline box first. The token lives in that browser's local storage and nowhere else; tap the chip again to sign out. The workflows behind the buttons see exactly the issue the link would have opened.
The per-paper buttons file GitHub issues that two workflows act on:
queue_actions.yml makes the mechanical moves (➕ ✅ ✖️ 🧹) and closes the
issue; queue_filing.yml has Claude file a 📥/📑 paper onto a
queue-filing/issue-<n> branch, gated, for a human to merge — and it opens
that PR itself when a QUEUE_PR_TOKEN repo secret (a fine-grained PAT with
pull-requests write on this repo) exists, or when Allow GitHub Actions to
create and approve pull requests is on (Settings → Actions → General). A tap
whose label the issue
form dropped is read from its title; whatever nothing acted on is
re-dispatched by the nightly queue_sweep.yml; and a filing that reached its
branch but not main sits at the top of the board under Filings awaiting
merge until its PR lands — the paper is not in memory before then.
🧠 162 pages · 71 drafted · 41% of 656 paper sections cite a resolved key · 228 papers queued · dashboard →
Self-contained, shared schema:
| Wiki | Covers |
|---|---|
wiki/lensing/ |
strong gravitational lensing (the primary wiki) |
wiki/smbh/ |
supermassive black holes, binaries, recoil, GW background |
wiki/cti/ |
charge transfer inefficiency, Euclid VIS calibration |
wiki/methods/ |
Bayesian inference, samplers, deep learning, simulations |
wiki/galaxies/ |
galaxy formation and evolution |
bibliography/ holds the canonical BibTeX
metadata every wiki cites against; reading-queue.md
is what's waiting to be read and filed. Two overnight tiers sit in front of it:
arxiv-inbox.md (strong lensing, a seven-day timer) and
arxiv-interests.md (everything else — black holes,
dark matter, galaxy formation, statistics — as one day's ten at a time, a
backlog you clear a day at a time rather than a timer). New knowledge updates
the metadata
and the claim support together, then passes make validate (CI-enforced on
every push).
Away for a while? /catch_up collects every
paper missed since memory last ingested one — lapsed suggestions from git
history, the open queue, and an arXiv gap-fill over days the digest never ran
(scripts/catch_up.py) — for one triage and one filing PR. The dashboard
shows a strong-lensing catch-up banner after seven days without recorded lensing
paper activity. Recent activity in other topics does not reset that clock.
The same lensing_catch_up record backs the banner and cockpit feed: canonical
state, cutoff date, age, seven-day threshold/deadline, observation time, evidence
links and explicit action descriptors. Healthy work adds no attention row;
missing, invalid or future dates are unknown. last_ingestion identifies the
verified bib-plus-sources commit, while last_completed is a queue DONE date
(which may mean read-without-filing); last_activity is the existing catch-up
cutoff, the later of those dates. Legacy all-scope fields remain available.
The cockpit copies an instruction to read skills/catch_up/SKILL.md and use
the catch_up skill with lensing, and links to its procedure. Its safety is scientific_judgement: candidate selection and filing
retain the procedure's human review. Observing staleness executes nothing, and
no scientific choice is fabricated before candidates are harvested.
The feed also exposes independent digests.lensing and digests.interests
records, shared with the board freshness indicators. Each records its canonical
state, last recorded date, checked time, elapsed weekdays, two-weekday threshold,
reason and evidence/action links. Empty queues with a recent stamp stay healthy;
missing, invalid or future stamps are unknown. The stamp proves a digest was
recorded, not that every workflow step succeeded. Stale/unknown cockpit rows
link to the owning Mind workflow and offer a manual investigation prompt
(requires_approval); detection never reruns a workflow. Digest delivery and
human lensing catch-up remain separate clocks.
The wiki schema is defined in
wiki/AGENTS.md and inherited by every
sub-wiki. How agents should read this repo: AGENTS.md. The
organism this repo is the Memory of:
PyAutoBrain/ORGANISM.md,
documented in full at https://pyautoscientist.readthedocs.io.
Licence: structure and tooling MIT; wiki content CC BY 4.0 — see LICENSE.