scrub: clean migration files, update stale numbers for launch

- migration/mem0_export: configurable collections via env var,
  remove hardcoded clawd/attractor/governor/gradient references
- migration/mem0_classify: generic entity examples, generic paths
- migration/mem0_generate: generic paths
- migration/run_mem0_backfill: generic path in log message
- CHANGELOG: VRAM -> memory budget
- ADR-001: 17 -> 18 MCP tools
- LAUNCH-POSTS-FINAL: 56 -> 92 tests, 17 -> 18 tools,
  updated HN post to final version

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: Paul Kyle

ADR-001-tools-over-pipeline.md

@@ -6,7 +6,7 @@
 
 ## Decision
 
-Palinode's primary value is its **17 MCP tools + file-based storage + git versioning**, not its 4-phase injection pipeline. The pipeline is scaffolding for current models. The tools survive any model upgrade.
+Palinode's primary value is its **18 MCP tools + file-based storage + git versioning**, not its 4-phase injection pipeline. The pipeline is scaffolding for current models. The tools survive any model upgrade.
 
 ## Context
 

docs/CHANGELOG.md

@@ -7,7 +7,7 @@
   - Commands: search, save, status, diff, consolidate, trigger, doctor
   - TTY-aware output (human text vs piped JSON)
   - Remote access via `PALINODE_API` env var or SSH
-- **GPU usage plane spec** — VRAM budget for embeddings + transcription + general LLM
+- **GPU usage plane spec** — memory budget for embeddings + transcription + general LLM
 - **Architecture decision: CLI vs MCP** — CLI for agents/scripts/cron (8x fewer tokens), MCP for IDEs only
 - Updated README architecture diagram to show CLI path
 - Updated ROADMAP with Phase 1.25 (CLI + interface rationalization)

docs/LAUNCH-POSTS-FINAL.md

@@ -13,7 +13,7 @@ Reply to: https://x.com/karpathy/status/2039805659525644595
 
 I've been building exactly this.
 
-Palinode: git-versioned markdown memory for AI agents. Hybrid BM25+vector search. 17 MCP tools. Deterministic compaction.
+Palinode: git-versioned markdown memory for AI agents. Hybrid BM25+vector search. 18 MCP tools. Deterministic compaction.
 
 The part I added? git blame on every fact your agent knows.
 
@@ -40,39 +40,35 @@ Same philosophy. With provenance.
 
 **Tweet 4 (with status screenshot):**
 
-227 files. 2,230 chunks indexed. 56 tests. 17 MCP tools. No cloud. No external DB. SQLite-vec + FTS5 + BGE-M3.
+227 files. 2,230 chunks indexed. 92 tests. 18 MCP tools. No cloud. No external DB. SQLite-vec + FTS5 + BGE-M3.
 
 Runs on a single box. MIT license.
 
 github.com/Paul-Kyle/palinode
 
 ---
 
-## Show HN (Monday ~8am ET / 5am PT)
+## Show HN (Wednesday ~9am PT)
 
 **Title:**
-Show HN: Palinode – Persistent agent memory as plain markdown with git provenance
+Show HN: Palinode – Git-versioned markdown memory for AI agents
 
-**Body:**
-Palinode is persistent memory for AI agents. Your agent's memory is a folder of markdown files — typed (people, projects, decisions, insights), git-versioned, and searchable with hybrid BM25 + vector.
+**Body (post as first comment):**
+I was using Mem0 for agent memory and found myself SSHing into a box to grep Qdrant vectors trying to figure out when my agent learned something wrong. I wanted `git blame` for agent memory. So I built it.
 
-The architecture is simple: markdown files are truth, SQLite (vec + FTS5) is a derived index, and every interface (MCP, REST API, CLI, agent plugin) hits the same backend. Set up on a server, connect from any IDE on any machine.
+Palinode stores your agent's memory as typed markdown files (people, projects, decisions, insights) with YAML frontmatter. A file watcher indexes them into SQLite-vec + FTS5 for hybrid search. Weekly, an LLM proposes structured compaction ops (KEEP/UPDATE/MERGE/SUPERSEDE/ARCHIVE) and a deterministic executor applies them — the LLM never writes files directly. Every compaction is a git commit.
 
-What made me build this:
+The part I think is actually new: diff, blame, rollback, and push are MCP tools your agent can call. Not just git-compatible files — the agent can trace any fact back to the session that recorded it, or revert a bad compaction, without you touching a terminal.
 
-I was using Mem0, then found myself grepping the Qdrant vectors trying to figure out when my agent learned something wrong. I wanted `git blame` for agent memory. So I built it.
+Architecture is dumb on purpose. Markdown files are truth. SQLite is a derived index. If everything crashes, `cat` still works. One API server, one .db file, one directory.
 
-What's different:
+It runs on a homelab box. I connect from two laptops over Tailscale. The MCP server is a pure HTTP client with no state — works with Claude Code, Cursor, Zed, anything that speaks MCP. Same 18 tools are also available as a REST API and CLI.
 
-- **Files are truth** — if every service crashes, `cat` still works. Rebuild the index anytime.
-- **Git operations as agent tools** — diff, blame, rollback, push are callable MCP tools, not just CLI conveniences.
-- **Deterministic compaction** — an LLM proposes structured ops (KEEP/UPDATE/MERGE/SUPERSEDE/ARCHIVE), a deterministic executor applies them and commits. The LLM never writes files directly.
-- **One backend, every interface** — MCP server (Streamable HTTP or stdio), REST API, CLI, OpenClaw plugin. Same 18 tools everywhere. Connect Claude Code, Zed, Cursor, Claude Desktop, or your own scripts.
-- **No infrastructure** — SQLite-vec + FTS5 + Ollama. No Postgres, no Redis, no cloud. One directory, one .db file, one API server.
+Karpathy's knowledge-base gist (https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f) articulated a lot of what I was already building toward — particularly the raw/compiled split, which maps directly to Palinode's ingest/consolidate cycle. This is one working implementation of those ideas.
 
-Stack: Python 3.11+, BGE-M3 embeddings via Ollama, any chat model for consolidation. 56 tests. MIT.
+What it doesn't do: no auto-injection into arbitrary LLM calls (you need an MCP client or to call the API). No multi-user. No cloud hosted version. It's a personal tool for one human and their agents.
 
-I run it on a homelab box and connect from two laptops over Tailscale. The MCP server is a pure HTTP client — it holds no state, just proxies to the API.
+Python 3.11+, BGE-M3 via Ollama, any chat model for consolidation. 92 tests. MIT.
 
 https://github.com/Paul-Kyle/palinode
 
@@ -100,7 +96,7 @@ It works from any IDE — the MCP server runs over Streamable HTTP, so Claude Co
 
 Browse your agent's brain in Obsidian. If Ollama dies, cat and grep still work.
 
-18 tools, 56 tests, MIT: https://github.com/Paul-Kyle/palinode
+18 tools, 92 tests, MIT: https://github.com/Paul-Kyle/palinode
 
 ---
 
@@ -116,6 +112,6 @@ The architecture: markdown files → SQLite-vec + FTS5 hybrid index → 4-phase
 
 The design bet: files are the source of truth, everything else is a derived index. One backend, multiple interfaces (MCP server over Streamable HTTP, REST API, CLI, OpenClaw plugin). 18 tools work identically across Claude Code, Zed, Cursor, or shell scripts.
 
-Behavioral spec in PROGRAM.md. 56 tests. MIT.
+Behavioral spec in PROGRAM.md. 92 tests. MIT.
 
 https://github.com/Paul-Kyle/palinode

palinode/migration/mem0_classify.py

@@ -11,14 +11,14 @@
     python -m palinode.migration.mem0_classify
 
 Input:
-    ~/clawd/palinode/migration/mem0_export.json
+    {PALINODE_DIR}/migration/mem0_export.json
 
 Output:
-    ~/clawd/palinode/migration/mem0_classified.json
+    {PALINODE_DIR}/migration/mem0_classified.json
 
 LLM Config:
-    URL: http://localhost:8080/v1 (Mac Studio Qwen 72B)
-    Model: /path/to/models/qwen25-72b-abliterated-mlx
+    URL: PALINODE_CLASSIFY_LLM_URL (default: http://localhost:8080)
+    Model: PALINODE_CLASSIFY_LLM_MODEL (any OpenAI-compatible chat model)
 """
 from __future__ import annotations
 
@@ -113,17 +113,17 @@ def deduplicate(memories: list[dict]) -> list[dict]:
    - Config: system configuration or technical setup detail
    - Skip: not worth keeping (too vague, trivially obvious, or stale operational detail)
 
-2. **entities**: Array of entity references, e.g. ["person/paul", "project/mm-kmd"]
-   Known entities: person/paul, person/peter, person/aidan, project/mm-kmd, project/palinode, project/color-class, project/infrastructure
+2. **entities**: Array of entity references, e.g. ["person/alice", "project/my-app"]
+   Detect entities from the content. Use the pattern: person/{name}, project/{slug}.
 
 3. **group**: A short slug for grouping related memories into the same file.
    Related memories about the same topic should share a group slug.
-   Examples: "mm-kmd-lora-training", "paul-preferences", "infrastructure-vllm"
+   Examples: "my-app-milestones", "alice-preferences", "infrastructure-setup"
 
 Return a JSON array with one object per memory:
 ```json
 [
-  {"index": 0, "type": "ProjectSnapshot", "entities": ["project/mm-kmd"], "group": "mm-kmd-milestones"},
+  {"index": 0, "type": "ProjectSnapshot", "entities": ["project/my-app"], "group": "my-app-milestones"},
   {"index": 1, "type": "Skip", "entities": [], "group": ""},
   ...
 ]

palinode/migration/mem0_export.py

@@ -8,7 +8,7 @@
     python -m palinode.migration.mem0_export
 
 Output:
-    ~/clawd/palinode/migration/mem0_export.json
+    {PALINODE_DIR}/migration/mem0_export.json
 """
 from __future__ import annotations
 
@@ -24,7 +24,10 @@
 logger = logging.getLogger("palinode.migration.mem0_export")
 
 QDRANT_URL = "http://localhost:6333"
-COLLECTIONS = ["mem0_attractor", "mem0_governor", "mem0_gradient"]
+COLLECTIONS = os.environ.get(
+    "PALINODE_MEM0_COLLECTIONS",
+    "mem0_default"
+).split(",")
 BATCH_SIZE = 100  # Qdrant scroll batch size
 
 
@@ -92,7 +95,7 @@ def _scroll_collection(collection: str) -> list[dict]:
                 "created_at": payload.get("createdAt", ""),
                 "hash": payload.get("hash", ""),
                 "session_type": payload.get("session_type", ""),
-                "user_id": payload.get("userId", "clawd"),
+                "user_id": payload.get("userId", "default"),
             })
 
         offset = result.get("next_page_offset")

palinode/migration/mem0_generate.py

@@ -8,10 +8,10 @@
     python -m palinode.migration.mem0_generate
 
 Input:
-    ~/clawd/palinode/migration/mem0_classified.json
+    {PALINODE_DIR}/migration/mem0_classified.json
 
 Output:
-    Markdown files in ~/clawd/palinode/{category}/
+    Markdown files in {PALINODE_DIR}/{category}/
 """
 from __future__ import annotations
 

palinode/migration/run_mem0_backfill.py

@@ -48,7 +48,7 @@ def main():
     logger.info(f"Total files: {sum(stats.values())}")
     logger.info("")
     logger.info("Next steps:")
-    logger.info("  1. Review the generated files in ~/clawd/palinode/")
+    logger.info("  1. Review the generated files in your PALINODE_DIR")
     logger.info("  2. Run 'curl -X POST http://localhost:6340/reindex' to index them")
     logger.info("  3. Test search: 'curl -X POST http://localhost:6340/search -d {\"query\":\"test\"}'")
     logger.info("  4. If satisfied, disable Mem0 autoRecall in OpenClaw config")