Public paper repository for:
Remember What Matters: State-Conditioned Episodic Retrieval for Emotional AI Companions
Author: Nikita Shilov, nik@zbs.gg
The paper studies a narrow retrieval primitive for emotional AI companions: the same user query may need different remembered episodes depending on the user's current emotional or contextual state.
For a quick non-specialist summary, read ONE_PAGER.md.
The current evidence is deliberately scoped:
- Pulse does not improve general retrieval overall on the primary n=100 deterministic suite.
- The supported result is stateful Recall@3 under a single-user deployment regression protocol: Pulse
0.419vs.cosine_state0.314andhybrid_state0.219. - The study is an engineering case study and regression artifact, not a cross-user benchmark or broad memory-system superiority claim.
This repository contains the public paper surface:
paper.tex
paper.pdf
ONE_PAGER.md
references.bib
REFERENCE_AUDIT.md
scripts/verify_bib_urls.py
docs/companion_memory_comparison_plan.md
PUBLICATION_NOTE.md
LICENSE
This public repository intentionally excludes:
- the private deployment corpus;
- raw or semi-raw companion logs;
- raw Apple Health exports;
- the actual Cartographer profile;
- raw judge notes and closed-review verdict artifacts;
- closed-review bundles and old arXiv/review zip/tar artifacts.
The closed-review verification packet can be shared privately with trusted reviewers. A public replication packet requires a separate redaction/privacy pass.
tectonic --keep-logs --keep-intermediates paper.texThe generated paper.pdf in this repository is built from paper.tex.
The public repository can verify bibliography hygiene without the private benchmark corpus:
make verify-refsThis checks cited arXiv, DOI, and URL targets and fails on missing references or title mismatches. REFERENCE_AUDIT.md records the last successful online reference audit.
The numerical claim verifier from the closed-review bundle is not included here because it depends on private/redacted benchmark artifacts.
The manuscript and accompanying text (paper.tex, paper.pdf, ONE_PAGER.md, references.bib, REFERENCE_AUDIT.md, docs/) are © 2026 Nikita Shilov and released under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0). See LICENSE. You may share and adapt the material for non-commercial purposes with attribution.
The helper code under scripts/ is additionally available under the MIT license.
The Pulse retrieval engine is licensed separately in its own repository: https://github.com/zbs-gg/pulse (AGPL-3.0, with commercial terms available).