diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
new file mode 100644
index 0000000..e943802
--- /dev/null
+++ b/.github/workflows/release.yml
@@ -0,0 +1,54 @@
+name: Release
+
+# Cuts a GitHub release (and its tag) from a chosen commit.
+# Run it from the Actions tab or via the API with:
+#   version: "0.4.0"           -> tag v0.4.0, must match pyproject.toml
+#   title:   "The Site Archive" (optional codename appended to the release name)
+#   target:  "main"            (branch or commit SHA the tag points at)
+# Release notes are read from docs/releases/v<version>.md.
+
+on:
+  workflow_dispatch:
+    inputs:
+      version:
+        description: "Version to release (e.g. 0.4.0)"
+        required: true
+      title:
+        description: "Optional codename, e.g. The Site Archive"
+        required: false
+        default: ""
+      target:
+        description: "Branch or commit SHA to tag"
+        required: false
+        default: "main"
+
+permissions:
+  contents: write
+
+jobs:
+  release:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          ref: ${{ inputs.target }}
+          fetch-depth: 0
+
+      - name: Check that pyproject.toml carries the version
+        run: |
+          grep -q "^version = \"${{ inputs.version }}\"" pyproject.toml \
+            || { echo "pyproject.toml does not declare version ${{ inputs.version }}"; exit 1; }
+          test -f "docs/releases/v${{ inputs.version }}.md" \
+            || { echo "docs/releases/v${{ inputs.version }}.md is missing"; exit 1; }
+
+      - name: Create tag and GitHub release
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          name="archaeocode v${{ inputs.version }}"
+          if [ -n "${{ inputs.title }}" ]; then name="$name — ${{ inputs.title }}"; fi
+          gh release create "v${{ inputs.version }}" \
+            --repo "${{ github.repository }}" \
+            --target "$(git rev-parse HEAD)" \
+            --title "$name" \
+            --notes-file "docs/releases/v${{ inputs.version }}.md"
diff --git a/docs/releases/v0.4.0.md b/docs/releases/v0.4.0.md
new file mode 100644
index 0000000..eadec1f
--- /dev/null
+++ b/docs/releases/v0.4.0.md
@@ -0,0 +1,35 @@
+**archaeocode** — automated software archaeology. Release 0.4.0, "The Site Archive".
+
+Until now every dig ended with a JSON report and amnesia. This release gives archaeocode a memory: a persistent, queryable **knowledge base** built on [LightRAG](https://github.com/HKUDS/LightRAG) (GraphRAG: knowledge graph + vector retrieval). Files, AST entities, dependency edges, and user stories become a graph you can interrogate from the terminal, from any MCP-capable agent, or as an Obsidian-style wiki.
+
+## Highlights
+
+- **Knowledge base workflow step** (`archaeo --knowledge-base`): files, classes, functions, methods, import/call edges with circular-dependency markers, and user stories are written into the graph **without a single LLM call**. Exact, cheap, idempotent.
+- **Business vocabulary on demand** (`--kb-extract`): with an API key, LightRAG's entity/relation extraction mines the code for customers, tariffs, invoice runs and files them next to the technical entities.
+- **`archaeo-kb` CLI**: `query`, `retrieve`, `context`, `graph`, `stats`, `add`, `wiki`. Only `query` needs an LLM; everything else works offline.
+- **`knowledge-base` MCP server** with `kb_query`, `kb_retrieve`, `kb_context`, `kb_graph`, `kb_stats`, `kb_add_documents`, `kb_export_wiki`, so Claude Desktop, Claude Code, or your own agent can ask the archive instead of re-reading the code.
+- **Obsidian-compatible wiki export** (`--kb-wiki DIR`): one cross-linked page per file, function, dependency, and user story.
+- **Scale by configuration**: NetworkX + nano-vectordb on a laptop with zero infrastructure; Neo4j / PostgreSQL+pgvector / Qdrant / Milvus / Redis for estates the size of a bank; workspaces keep several systems apart in one back-end. Embeddings via sentence-transformers (local, default), OpenAI, or Ollama; LLM via Anthropic, OpenAI, or Ollama.
+- **Air-gapped mode**: `KB_TOKENIZER=bytes` and `KB_EMBEDDING_PROVIDER=hashing` run the whole pipeline with no network. The new test suite runs this way in CI.
+- **Docs, retold as a dig**: README field guide, quickstart, and the [knowledge-management decision record](https://github.com/osick/archaeocode/blob/main/docs/KNOWLEDGE_MANAGEMENT.md) explaining why GraphRAG over an LLM wiki or vanilla RAG, and why LightRAG over Microsoft GraphRAG, Cognee, Graphiti, and neo4j-graphrag.
+
+## Getting started
+
+```bash
+git clone https://github.com/osick/archaeocode.git
+cd archaeocode
+pip install -r requirements.txt
+pip install -e .            # puts archaeo and archaeo-kb on your PATH
+cp .env.example .env        # add ANTHROPIC_API_KEY for user stories and answers
+
+archaeo --source sample_data/cobol --source-lang cobol --knowledge-base --kb-wiki ./wiki
+archaeo-kb retrieve "PAYMENT" --mode local
+archaeo-kb query "Which programs touch the customer master file?"
+```
+
+## Changed
+
+- AST entities now carry `file_path`, so classes, functions, and methods link to their file.
+- `requirements.txt` adds `lightrag-hku`.
+
+Full change history: [docs/CHANGELOG.md](https://github.com/osick/archaeocode/blob/main/docs/CHANGELOG.md)