docs: add paper PDF, branch model, dev CI trigger

.github/workflows/ci.yml

@@ -2,7 +2,7 @@ name: ci
 
 on:
   push:
-    branches: [main]
+    branches: [main, dev]
   pull_request:
   workflow_dispatch:
 

CONTRIBUTING.md

@@ -102,6 +102,22 @@ Code lives in `bench/forgeteval/generate.py`
 (or `lang_zh.py` / `lang_ja.py` for non-English).  Update the
 sub-template table in the docs in the same PR.
 
+## Branch model
+
+We use a three-tier flow:
+
+- `main` — release-quality; protected.  Force-pushes and deletions are
+  blocked; all five CI checks (pytest × py3.10–3.13 + ruff) must pass
+  before merge.  Only mergeable from `dev` via PR.
+- `dev` — integration; protected with the same rules as `main`.  All
+  feature work lands here first.
+- `feat/<short-slug>` — feature working branches.  No GitHub
+  protection; rebase, force-push, and squash freely.  Open a PR into
+  `dev` when ready; CI runs against the PR.
+
+Hotfixes can go directly to `main` via PR if the change is
+release-blocking and small.
+
 ## Pull request guidance
 
 - **One logical change per PR.**  Two unrelated improvements should

README.md

@@ -266,6 +266,24 @@ Roadmap (next):
 
 <br />
 
+## Paper
+
+📄 **[ForgetEval: Benchmarking the Forgetting Axis of Agent Memory Systems](paper/paper.pdf)** — Dongxu Yang, *DeepLethe*, 2026.
+
+Full methodology, formal model of the depth axis (Propositions 1–4), 1000-case ForgetEval, 5-seed variance, distractor sweep, component ablations, and LongMemEval-S comparison.  23 pages, MIT.
+
+```bibtex
+@misc{yang2026forgeteval,
+  author       = {Yang, Dongxu},
+  title        = {ForgetEval: Benchmarking the Forgetting Axis of
+                   Agent Memory Systems},
+  year         = {2026},
+  howpublished = {\url{https://github.com/deeplethe/lethe/blob/main/paper/paper.pdf}},
+}
+```
+
+<br />
+
 ## Star History
 
 <a href="https://star-history.com/#deeplethe/lethe&Date">