The manuscript is BrainBox: Local Behavioral Memory for Coding Agents, version 2 (8 October 2026), published on Zenodo as doi:10.5281/zenodo.23250505. It is a new version under concept DOI 10.5281/zenodo.18664906, which resolves to the latest version. The published file is the brainbox.pdf built from commit 6dc1ae9 (MD5 5c2c881d4f8fbe33e62074f348000fc8).
../WHITEPAPER.md: manuscript source, with Pandoc metadata, equations and citation keys. Figures point atassets/*.svgso they display on GitHub.references.bib: 16 references, each checked against its primary source.figures.mjs: exports four README figures (the loop, Hebbian learning, spreading activation, results) as vector PDFs in their settled frame, without the plate's own "fig. N" line. Output goes tofigures/(ignored).brainbox.tex: generated standalone LaTeX, including the resolved bibliography. Regenerate rather than editing it independently.brainbox.pdf: compiled reading copy (ignored by git; upload this file to Zenodo).evidence.json: aggregate data from four existing evaluation artifacts, their original paths and SHA-256 hashes. No private session text, task IDs or local project paths are included.build.sh: exports the figures, then regenerates LaTeX and PDF from the Markdown source.gemini-review.md: historical external critique; not a verified reference or an evaluation artifact. Some statements describe obsolete code, so the revised manuscript is grounded in current code and raw results instead.
Requires Node, Chromium (for the figure export), Pandoc with citeproc, and Tectonic. This version was compiled with Pandoc 3.9 and Tectonic 0.17.0; Tectonic downloads its TeX bundle, including the TeX Gyre Pagella text and math fonts, on the first build.
# From the repository root:
bash paper/build.sh
# Or choose explicit binary paths:
PANDOC=/path/to/pandoc TECTONIC=/path/to/tectonic CHROMIUM=/path/to/chromium bash paper/build.shIf a figure scene changes, run npm run visuals first so assets/*.svg is current. No tests, model downloads or install hooks are involved.
The revision replaces the older draft's headline estimated-token-savings claims with coverage-conditioned replay results. It does not report a new benchmark run. Numerical tables were checked against the original JSON artifacts, not copied from the README or scout summaries.
- Public cohort: 40 selected trajectories; only 16 edited-file tasks, 30 read-file tasks, 221 next-file opportunities and 2,625 next-tool opportunities are scored after prior-history eligibility filtering.
- Local cohort: 97 selected sessions from up to five projects; metadata's nine projects refers to the full scanned source corpus.
- File scores are MRR@10; tool scores are MRR@3. Differences and intervals are against the named heuristic, not against the historical BrainBox engine.
- Current-run artifacts lack source commit hashes and selected input manifests. The inspected source revision is recorded separately and is not asserted to be the exact executable that produced them.
- The public historical/current runs match declared options, coverage and comparator scores. Local source scanning changed between historical and current runs, so the draft does not claim a controlled local version comparison.
- Embeddings are disabled in the main runs. Injection-policy benefit, procedure quality, indexed code-search benefit, actual token savings and end-to-end task success remain unmeasured here.
evidence.json allows aggregate-table checking, but cannot reconstruct raw trajectories, re-estimate intervals from individual probes or establish missing original-run provenance. Original eval-results/ directories are ignored; fresh clones must rerun the public evaluation or obtain the underlying artifacts separately.
- Rewrote the paper around one question (how well an agent's own history predicts what it touches next) and the v0.2 implementation.
- Added four vector figures. The Hebbian figure was corrected to the update rule's actual values (0.10, 0.38, 0.66 after 1, 5 and 20 sessions; ceiling 1.0); the README and website copies were corrected with it.
- Added coverage denominators, all five baselines, paired intervals, ordering assumptions and provenance limits. Every number in Tables 2–4 and Appendix A was checked against
evidence.json. - Added Section 7, which lists each version-1 claim and its status: withdrawn, superseded, reframed or removed.
- Corrected related-work metadata and added three foundational references (Hebb 1949, Collins & Loftus 1975, Palmer & Zdonik 1991).
- Prepared with coding-assistant drafting and evidence review.