@bcmyguest/openmem 0.0.0-stage → 0.8.0

package/indexer/sinks/mem0.py

@@ -0,0 +1,247 @@
+"""mem0 sink: the self-hosted mem0 REST server (`server/main.py` in github.com/mem0ai/mem0), stdlib only.
+
+Namespace mapping. `bank` is the mem0 `user_id` and `doc_id` is the mem0 `run_id`, both through `scope_id`. mem0
+trims an entity id and then rejects one that still contains whitespace (`_validate_and_trim_entity_id` in
+`mem0/memory/main.py`; it has no length or charset rule), so `scope_id` is the identity for any id without
+whitespace, and otherwise replaces each run of whitespace with `_` and appends `~` plus the first 8 hex digits of the
+SHA-256 of the original id (`obsidian::My Vault` -> `obsidian::My_Vault~<8 hex>`). The suffix keeps two different
+ids that differ only in whitespace apart, and the mapping is deterministic, so the same bank and document always
+resolve to the same ids. It is applied to every use of the two ids (the lookup, the writes, the deletes). The
+original text of a mapped id is not stored. Every memory also carries the fixed `agent_id`
+`openmem-indexer`, for provenance only: lookups filter on `user_id` and `run_id` and never on `agent_id`.
+mem0 has no bank to create (the `user_id` exists once a memory carries it), so `prepare` is a no-op.
+
+Documents. mem0 gives every memory a server-generated id and has no upsert, so the sink resolves a document to its
+memories with `GET /memories?user_id=<bank>&run_id=<doc_id>&top_k=1000` (percent-encoded). `retain` is
+write-then-delete so that a failure never loses a document: it lists the document's current memory ids, writes the
+new text, and only then deletes the listed ids (`DELETE /memories/{id}`; a 404 for an already-gone id is swallowed).
+A failure part-way leaves the old memories, possibly next to some of the new ones; the indexer keeps the document
+marked as changed, and its next run repeats the whole sequence, which converges on exactly the new chunks. Until
+then a search can see duplicates. The text is split into chunks of at most 3000 characters (cut at a blank line,
+else a newline, else a space, else mid-text; the chunks concatenate back to the original text, only chunks that are
+pure whitespace are dropped), and each chunk is one `POST /memories` with `infer` literally false, so mem0 stores it
+verbatim instead of asking an LLM to rewrite it. Each reply must be a JSON object whose `results` list is not empty
+(infer=false answers one ADD result per message), otherwise the retain fails. A document of empty or whitespace-only
+text has no chunks: its old memories are deleted and nothing is written (the document is simply absent, which is
+what an empty document means to a search).
+
+Limits and races. A document that would need more than 900 chunks (about 2.7 million characters) is rejected before
+any request is sent, with a message naming the document id, so the 1000-entry listing can never hide a stale chunk
+and a rejected retain leaves the previous version intact. The indexer retries a failed document on every run, so such
+a document fails its whole source on every run until it is shortened or excluded. mem0 enforces no uniqueness on
+`run_id`: two indexer runs that retain the same document concurrently (for example with different state files) can
+both list the old memories and both write, leaving two sets of chunks. The next retain of that document lists and
+removes all of them.
+
+Metadata. The caller's metadata is copied; a key that mem0 reserves, drops or overwrites in the stored payload
+(`data`, `user_id`, `agent_id`, `run_id`, `actor_id`, `hash`, `created_at`, `updated_at`, `expiration_date`, `role`,
+`text_lemmatized`, `attributed_to`) is renamed with the prefix `meta_`. The sink then adds `tags` (a list, only when
+there are tags), `context` and `timestamp` (only when not None; mem0 has no event-time field) and, for a document of
+several chunks, `chunk` (0-based index) and `chunks` (the total). These sink-owned keys win over a caller key of
+the same name.
+
+Auth. `MEM0_API_KEY` is read from the environment inside `create`, never from a document or a path, and is sent as
+`X-API-Key` on every request except the health check (`/docs` is open, so no key travels with it). Without a key no
+auth header is sent. The key never appears in an error message. Over plain `http://` the key travels in cleartext:
+use `https://` (or a loopback address) for any server that is not on the same host. The sink follows no redirects
+(see `sinks/_http.py`; the mem0 server has `redirect_slashes=False` and API calls never redirect), so a 3xx is an
+error carrying its status, and the key cannot be handed to another host.
+
+URL. There is no default: mem0 publishes no standard port. `create` reads `MEMORY_API_URL` when no url is passed and
+raises `MissingUrlError` (a `SinkError` and a `ValueError`, so the CLI reports it as a sink error) when both are empty,
+when the URL has no http(s) scheme or host, or when it embeds credentials (`user:pass@host`; use `MEM0_API_KEY`).
+Error messages never echo the URL given. Health is `GET /docs`, which the server leaves open (it has no dedicated
+health route).
+"""
+
+import hashlib
+import json
+import os
+import urllib.error
+import urllib.parse
+import urllib.request
+from collections.abc import Mapping
+
+from sinks import Capabilities, SinkError
+from sinks._http import open_url, read_all
+
+URL_ENV = "MEMORY_API_URL"
+KEY_ENV = "MEM0_API_KEY"
+AGENT_ID = "openmem-indexer"
+CHUNK_CHARS = 3000
+MAX_CHUNKS = 900
+LIST_LIMIT = 1000
+RESERVED_KEYS = frozenset(
+    {"data", "user_id", "agent_id", "run_id", "actor_id", "hash", "created_at", "updated_at", "expiration_date", "role", "text_lemmatized", "attributed_to"}
+)
+READ_TIMEOUT = 30
+WRITE_TIMEOUT = 300
+HEALTH_TIMEOUT = 5
+
+
+class MissingUrlError(SinkError, ValueError):
+    """No usable API URL. A ValueError too, so `bin/hindsight-project-index` reports it as a sink error (exit 2)."""
+
+
+class HttpError(SinkError):
+    def __init__(self, method: str, url: str, status: int, body: str):
+        super().__init__(f"{method} {url}: HTTP {status}: {body}", status)
+
+
+def scope_id(value: str) -> str:
+    """Map a bank or document id to a mem0 entity id: itself when it has no whitespace, else a whitespace-free form.
+
+    Each run of whitespace becomes `_`, then `~` and the first 8 hex digits of sha256(value) are appended.
+    """
+    if not any(char.isspace() for char in value):
+        return value
+    pieces = []
+    in_space = False
+    for char in value:
+        if char.isspace():
+            if not in_space:
+                pieces.append("_")
+            in_space = True
+        else:
+            pieces.append(char)
+            in_space = False
+    return "".join(pieces) + "~" + hashlib.sha256(value.encode()).hexdigest()[:8]
+
+
+def split_chunks(text: str, limit: int = CHUNK_CHARS) -> list[str]:
+    """Cut `text` into pieces of at most `limit` characters whose concatenation is `text`, minus pure-whitespace pieces."""
+    pieces = []
+    position = 0
+    while len(text) - position > limit:
+        window = text[position : position + limit]
+        cut = limit
+        for separator in ("\n\n", "\n", " "):
+            found = window.rfind(separator)
+            if found >= 0:
+                cut = found + len(separator)
+                break
+        pieces.append(text[position : position + cut])
+        position += cut
+    pieces.append(text[position:])
+    return [piece for piece in pieces if piece.strip()]
+
+
+class Mem0Sink:
+    capabilities = Capabilities(async_ingest=False)
+
+    def __init__(self, url: str, api_key: str | None = None):
+        self.url = url.rstrip("/")
+        self._api_key = api_key or None
+
+    def _headers(self, with_body: bool) -> dict[str, str]:
+        headers = {"Content-Type": "application/json"} if with_body else {}
+        if self._api_key:
+            headers["X-API-Key"] = self._api_key
+        return headers
+
+    def _excerpt(self, body: bytes) -> str:
+        text = body.decode(errors="replace")[:300]
+        return text.replace(self._api_key, "***") if self._api_key else text
+
+    def _request(self, method: str, path: str, payload: dict | None = None, timeout: float = READ_TIMEOUT) -> object:
+        """One request; returns the decoded JSON reply, or None for an empty or non-JSON body."""
+        url = self.url + path
+        data = json.dumps(payload).encode() if payload is not None else None
+        try:
+            with open_url(urllib.request.Request(url, data=data, headers=self._headers(data is not None), method=method), timeout) as response:
+                body = read_all(response)
+        except urllib.error.HTTPError as error:
+            try:
+                excerpt = self._excerpt(read_all(error))
+            except OSError:
+                excerpt = ""
+            raise HttpError(method, url, error.code, excerpt) from error
+        try:
+            return json.loads(body) if body else None
+        except ValueError:
+            return None
+
+    def _memory_ids(self, user_id: str, run_id: str) -> list[str]:
+        query = urllib.parse.urlencode({"user_id": user_id, "run_id": run_id, "top_k": LIST_LIMIT}, quote_via=urllib.parse.quote)
+        reply = self._request("GET", "/memories?" + query)
+        results = reply.get("results") if isinstance(reply, dict) else None
+        if not isinstance(results, list) or not all(isinstance(item, dict) and isinstance(item.get("id"), str) and item["id"] for item in results):
+            raise SinkError(f"GET {self.url}/memories: unexpected reply (want an object with a `results` list of memories that have an id); cannot tell which memories the document has")
+        return [item["id"] for item in results]
+
+    def _delete_ids(self, memory_ids: list[str]) -> None:
+        for memory_id in memory_ids:
+            try:
+                self._request("DELETE", "/memories/" + urllib.parse.quote(memory_id, safe=""))
+            except HttpError as error:
+                # 404: already gone (deleted server-side or by a concurrent run).
+                if error.status != 404:
+                    raise
+
+    def health(self) -> bool:
+        try:
+            with open_url(urllib.request.Request(self.url + "/docs"), HEALTH_TIMEOUT) as response:
+                return 200 <= response.status < 300
+        except (OSError, ValueError):  # includes HTTPError, URLError and, via open_url, malformed HTTP replies
+            return False
+
+    def prepare(self, bank: str) -> None:
+        return None
+
+    def retain(
+        self,
+        bank: str,
+        doc_id: str,
+        content: str,
+        *,
+        tags: list[str] | None = None,
+        metadata: Mapping[str, str] | None = None,
+        context: str | None = None,
+        timestamp: str | None = None,
+    ) -> None:
+        chunks = split_chunks(content)
+        if len(chunks) > MAX_CHUNKS:
+            raise SinkError(f"document {doc_id!r} needs {len(chunks)} chunks of at most {CHUNK_CHARS} characters; the mem0 sink accepts at most {MAX_CHUNKS} per document")
+        user_id, run_id = scope_id(bank), scope_id(doc_id)
+        base = {f"meta_{key}" if key in RESERVED_KEYS else key: value for key, value in (metadata or {}).items()}
+        if tags:
+            base["tags"] = list(tags)
+        if context is not None:
+            base["context"] = context
+        if timestamp is not None:
+            base["timestamp"] = timestamp
+        old_ids = self._memory_ids(user_id, run_id)
+        for index, chunk in enumerate(chunks):
+            meta = dict(base)
+            if len(chunks) > 1:
+                meta["chunk"] = index
+                meta["chunks"] = len(chunks)
+            body = {"messages": [{"role": "user", "content": chunk}], "user_id": user_id, "agent_id": AGENT_ID, "run_id": run_id, "infer": False, "metadata": meta}
+            reply = self._request("POST", "/memories", body, WRITE_TIMEOUT)
+            results = reply.get("results") if isinstance(reply, dict) else None
+            if not isinstance(results, list) or not results:
+                raise SinkError(f"POST {self.url}/memories: mem0 stored nothing for chunk {index + 1} of {len(chunks)} of document {doc_id!r} (want an object with a non-empty `results` list)")
+        self._delete_ids(old_ids)
+
+    def delete(self, bank: str, doc_id: str) -> None:
+        self._delete_ids(self._memory_ids(scope_id(bank), scope_id(doc_id)))
+
+
+def create(url: str | None = None, **_cfg) -> Mem0Sink:
+    resolved = (url or os.environ.get(URL_ENV) or "").strip()
+    if not resolved:
+        raise MissingUrlError(f"mem0 sink needs an API URL: set {URL_ENV} or pass --url")
+    # No message below echoes the URL: it could carry a password.
+    if not resolved.lower().startswith(("http://", "https://")):
+        raise MissingUrlError("mem0 sink API URL must start with http:// or https://")
+    try:
+        parts = urllib.parse.urlsplit(resolved)
+        host = parts.hostname
+        _ = parts.port  # raises ValueError for a malformed port
+    except ValueError:
+        raise MissingUrlError("mem0 sink API URL is not a valid URL") from None
+    if "@" in parts.netloc:
+        raise MissingUrlError(f"mem0 sink API URL must not contain credentials (user:pass@host); set {KEY_ENV} for the API key")
+    if not host:
+        raise MissingUrlError("mem0 sink API URL has no host")
+    return Mem0Sink(resolved, (os.environ.get(KEY_ENV) or "").strip() or None)

package/indexer/watch.py

@@ -0,0 +1,240 @@
+#!/usr/bin/env python3
+"""Keep Hindsight current: watch Obsidian vaults with inotify, poll git repositories, re-list sources periodically.
+
+inotify is not used for repositories: they hold hundreds of thousands of directories, which exceeds the host's
+max_user_watches limit. Repositories are polled by fingerprint (HEAD plus tracked-file status) instead.
+"""
+
+import hashlib
+import os
+import subprocess
+import sys
+import threading
+import time
+from collections.abc import Mapping
+from dataclasses import dataclass
+from pathlib import Path
+
+import index_config
+
+
+INDEXER = "hindsight-project-index"
+MAX_BACKOFF_SECONDS = 600
+IGNORED_DIRS = {"node_modules", "build", "dist", "target", "vendor", "__pycache__"}
+DRY_RUN = os.environ.get("INDEXER_DRY_RUN") == "1"
+
+
+def timings(settings: index_config.Config, environ: Mapping[str, str]) -> tuple[int, int, int]:
+    """Return (poll, vault quiet, relist) seconds. Environment variables override index.yml."""
+    return (
+        int(environ.get("POLL_SECONDS", settings.poll_seconds)),
+        int(environ.get("VAULT_QUIET_SECONDS", settings.vault_quiet_seconds)),
+        int(environ.get("RELIST_SECONDS", settings.relist_seconds)),
+    )
+
+
+# Loaded at import, so an invalid config stops the watcher before it starts.
+POLL_SECONDS, VAULT_QUIET_SECONDS, LIST_SECONDS = timings(index_config.load(), os.environ)
+
+
+@dataclass(frozen=True)
+class RepoWatch:
+    indexed: str | None  # fingerprint at the last successful index
+    seen: str | None  # fingerprint at the previous poll
+
+
+def log(message: str) -> None:
+    print(f"{time.strftime('%Y-%m-%d %H:%M:%S')} {message}", flush=True)
+
+
+def git_fingerprint(repo: Path) -> str:
+    """HEAD plus a digest of tracked-file changes. Untracked files are ignored, matching the indexer."""
+    head = subprocess.run(["git", "rev-parse", "--verify", "-q", "HEAD"], cwd=repo, capture_output=True, text=True, check=False)
+    head_id = head.stdout.strip() if head.returncode == 0 else "unborn"
+    status = subprocess.run(["git", "status", "--porcelain=v1", "-uno"], cwd=repo, capture_output=True, check=True).stdout
+    return f"{head_id}:{hashlib.sha256(status).hexdigest()}"
+
+
+def step(state: RepoWatch, fingerprint: str) -> tuple[RepoWatch, bool]:
+    """Advance one poll. Return the new state and whether to index now.
+
+    A change indexes only after the fingerprint stays unchanged for one more poll.
+    """
+    new_state = RepoWatch(state.indexed, fingerprint)
+    if fingerprint == state.indexed:
+        return new_state, False
+    return new_state, fingerprint == state.seen
+
+
+def mark_indexed(state: RepoWatch, fingerprint: str) -> RepoWatch:
+    return RepoWatch(fingerprint, state.seen)
+
+
+def list_sources() -> dict[Path, str]:
+    result = subprocess.run([INDEXER, "--list"], capture_output=True, text=True, check=True)
+    sources: dict[Path, str] = {}
+    for line in result.stdout.splitlines():
+        kind, _, path = line.partition("\t")
+        if kind and path:
+            sources[Path(path)] = kind
+    return sources
+
+
+def root_for(path: Path, roots: list[Path]) -> Path | None:
+    matches = [root for root in roots if path == root or root in path.parents]
+    return max(matches, key=lambda root: len(root.parts), default=None)
+
+
+def make_filter(roots: list[Path]):
+    def keep(_change, raw_path: str) -> bool:
+        path = Path(raw_path)
+        root = root_for(path, roots)
+        if root is None:
+            return False
+        # Dot-prefixed paths cover .git, .obsidian, .trash, .venv, and dotfile caches.
+        return not any(part.startswith(".") or part in IGNORED_DIRS for part in path.relative_to(root).parts)
+
+    return keep
+
+
+def run_indexer(kind: str, path: Path) -> int:
+    flag = {
+        "repo": "--repo",
+        "vault": "--vault",
+        "notes": "--notes",
+        "papers": "--papers",
+        "bookmarks": "--bookmarks",
+        "feed": "--feed",
+    }[kind]
+    command = [INDEXER, "--apply", flag, str(path)]
+    log(f"index {kind} {path}")
+    if DRY_RUN:
+        log(f"dry run, not executed: {' '.join(command)}")
+        return 0
+    try:
+        code = subprocess.run(command, check=False).returncode
+    except OSError as error:
+        log(f"index {kind} {path} could not start: {error}")
+        return 1
+    log(f"index {kind} {path} exited {code}")
+    return code
+
+
+def watch_vaults(vaults: list[Path], stop: threading.Event) -> None:
+    """Index each vault after VAULT_QUIET_SECONDS without changes. Returns when stop is set."""
+    from watchfiles import watch  # imported here so the pure helpers work without watchfiles
+
+    pending: dict[Path, float] = {}
+    for changes in watch(*vaults, watch_filter=make_filter(vaults), rust_timeout=1000, yield_on_timeout=True, stop_event=stop):
+        for _change, raw_path in changes:
+            root = root_for(Path(raw_path), vaults)
+            if root is not None:
+                pending[root] = time.monotonic()
+        for root, last in list(pending.items()):
+            if time.monotonic() - last >= VAULT_QUIET_SECONDS:
+                del pending[root]
+                run_indexer("vault", root)
+                # Changes made during the run arrive in later batches and re-trigger the vault.
+
+
+class VaultWatcher:
+    def __init__(self, vaults: list[Path]):
+        self.vaults = vaults
+        self.stop = threading.Event()
+        self.thread = threading.Thread(target=self._run, name="vault-watch", daemon=True)
+
+    def start(self) -> "VaultWatcher":
+        self.thread.start()
+        return self
+
+    def close(self) -> None:
+        # Does not join: a running index may take minutes, and the poller must keep going.
+        self.stop.set()
+
+    def _run(self) -> None:
+        backoff = 1
+        while not self.stop.is_set():
+            try:
+                watch_vaults(self.vaults, self.stop)
+                return
+            except Exception as error:
+                log(f"vault watcher failed: {type(error).__name__}: {error}; retrying in {backoff}s")
+                self.stop.wait(backoff)
+                backoff = min(backoff * 2, MAX_BACKOFF_SECONDS)
+
+
+def index_repo(path: Path, repos: dict[Path, RepoWatch]) -> None:
+    try:
+        baseline: str | None = git_fingerprint(path)
+    except Exception as error:
+        log(f"{path}: fingerprint failed: {type(error).__name__}: {error}")
+        baseline = None
+    indexed = baseline if run_indexer("repo", path) == 0 else None
+    repos[path] = RepoWatch(indexed, baseline)
+
+
+def poll_repos(repos: dict[Path, RepoWatch]) -> None:
+    for path, state in list(repos.items()):
+        try:
+            fingerprint = git_fingerprint(path)
+        except Exception as error:
+            log(f"{path}: fingerprint failed: {type(error).__name__}: {error}")
+            continue
+        repos[path], should_index = step(state, fingerprint)
+        if should_index and run_indexer("repo", path) == 0:
+            repos[path] = mark_indexed(repos[path], fingerprint)
+
+
+def refresh(kinds: dict[Path, str], repos: dict[Path, RepoWatch], watcher: VaultWatcher | None) -> tuple[dict[Path, str], VaultWatcher | None]:
+    try:
+        found = list_sources()
+    except Exception as error:
+        log(f"listing sources failed, keeping {len(kinds)} known: {type(error).__name__}: {error}")
+        return kinds, watcher
+    if not found:
+        log("no sources selected; nothing to index or watch")
+    for path in set(kinds) - set(found):
+        log(f"source removed: {path}")
+        repos.pop(path, None)
+    watch_kinds = ("vault", "notes", "papers", "bookmarks", "feed")
+    watched = sorted(path for path, kind in found.items() if kind in watch_kinds)
+    if watched != sorted(path for path, kind in kinds.items() if kind in watch_kinds):
+        # Restart the watcher before indexing new vaults, so edits made meanwhile are queued, not lost.
+        if watcher is not None:
+            watcher.close()
+        watcher = VaultWatcher(watched).start() if watched else None
+    for path, kind in found.items():
+        if path in kinds:
+            continue
+        log(f"new {kind} source: {path}")
+        if kind in watch_kinds:
+            run_indexer(kind, path)
+        else:
+            index_repo(path, repos)
+    return found, watcher
+
+
+def main() -> int:
+    log(f"poll {POLL_SECONDS}s, vault quiet {VAULT_QUIET_SECONDS}s, source listing {LIST_SECONDS}s" + (", dry run" if DRY_RUN else ""))
+    kinds: dict[Path, str] = {}
+    repos: dict[Path, RepoWatch] = {}
+    watcher: VaultWatcher | None = None
+    next_list = 0.0
+    failures = 0
+    while True:
+        try:
+            if time.monotonic() >= next_list:
+                next_list = time.monotonic() + LIST_SECONDS
+                kinds, watcher = refresh(kinds, repos, watcher)
+            poll_repos(repos)
+            failures = 0
+            time.sleep(POLL_SECONDS)
+        except Exception as error:
+            failures += 1
+            delay = min(POLL_SECONDS * 2**failures, MAX_BACKOFF_SECONDS)
+            log(f"iteration failed: {type(error).__name__}: {error}; retrying in {delay}s")
+            time.sleep(delay)
+
+
+if __name__ == "__main__":
+    sys.exit(main())

package/package.json

@@ -1,6 +1,98 @@
 {
   "name": "@bcmyguest/openmem",
-  "version": "0.0.0-stage",
-  "stub": true,
-  "description": "Temporary package placeholder for staged publishing"
-}
+  "version": "0.8.0",
+  "description": "OpenMem — memory interface for coding agents: pluggable providers (Hindsight, mem0, Cognee, …), shared wizard/indexer, multi-format import/export (PMM and more).",
+  "type": "module",
+  "license": "Apache-2.0",
+  "author": "B Guest",
+  "homepage": "https://bcmyguest.github.io/openmem/",
+  "bugs": {
+    "url": "https://github.com/bcmyguest/openmem/issues"
+  },
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/bcmyguest/openmem.git"
+  },
+  "mcpName": "io.github.bcmyguest/openmem",
+  "keywords": [
+    "openmem",
+    "memory",
+    "ai-memory",
+    "hindsight",
+    "mem0",
+    "cognee",
+    "mcp",
+    "mcp-server",
+    "docker-compose",
+    "setup-wizard",
+    "coding-agent",
+    "pmm"
+  ],
+  "bin": {
+    "openmem": "cli/openmem.mjs",
+    "openmem-setup": "wizard/bin/setup.mjs",
+    "openmem-mcp": "cli/openmem-mcp.mjs",
+    "openmem-install-mcp": "cli/install-memory-mcp.mjs",
+    "openmem-install-coding-agent": "cli/install-coding-agent.mjs",
+    "openmem-switch-provider": "cli/switch-memory-provider.mjs",
+    "openmem-notebooklm-mcp": "cli/notebooklm-mcp.mjs",
+    "openmem-cognee-mcp": "cli/cognee-mcp.mjs"
+  },
+  "files": [
+    "cli/",
+    ".agents/skills/openmem/",
+    "wizard/bin/",
+    "wizard/lib/",
+    "providers/",
+    "compose.yaml",
+    "compose.bench.yaml",
+    "indexer/",
+    "bin/",
+    "sync/",
+    "proxy/",
+    "config/*.example",
+    "config/index.yml",
+    "config/notebooklm.yml",
+    "config/litellm.*.yaml",
+    "config/coding-agent.json",
+    "scripts/*.mjs",
+    "scripts/*.sh",
+    "README.md",
+    "LICENSE",
+    "glama.json",
+    "server.json",
+    "docs/",
+    "!docs/ROADMAP*",
+    "!docs/WAVE*",
+    "!docs/ANNOUNCEMENT*",
+    "!docs/PUBLISH*",
+    "!**/__pycache__",
+    "!**/*.pyc",
+    "!**/tests",
+    "!**/.pytest_cache"
+  ],
+  "engines": {
+    "node": ">=24.0.0"
+  },
+  "publishConfig": {
+    "access": "public"
+  },
+  "scripts": {
+    "setup": "node wizard/bin/setup.mjs",
+    "pm": "node wizard/bin/pm.mjs",
+    "openmem": "node cli/openmem.mjs",
+    "install-mcp": "node cli/install-memory-mcp.mjs",
+    "install-coding-agent": "node cli/install-coding-agent.mjs",
+    "feed-sync": "node cli/feed-sync.mjs",
+    "switch-provider": "node cli/switch-memory-provider.mjs",
+    "test": "for t in scripts/tests/test_*.mjs; do node \"$t\" || exit 1; done && bash scripts/tests/test_wizard.sh",
+    "test:indexer": "cd indexer && python3 -m pytest -q",
+    "test:unit": "node scripts/tests/test_cli_package.mjs --unit",
+    "pack:check": "node scripts/check-pack.mjs"
+  },
+  "dependencies": {
+    "@clack/prompts": "^1.2.0",
+    "@iarna/toml": "^2.2.5",
+    "yaml": "^2.9.1"
+  }
+}

package/providers/_template/compose.yaml

@@ -0,0 +1,52 @@
+# TEMPLATE compose fragment for a memory provider. Copy to providers/<id>/compose.yaml and replace every
+# "example" / EXAMPLE_ below with the provider id. Directories starting with "_" are not providers; the wizard's
+# provider registry must skip them.
+#
+# Layering (the wizard passes both, then `up -d`):
+#   docker compose -f compose.yaml -f providers/<id>/compose.yaml --profile <id> up -d
+# Add `-f compose.override.yaml` when you use the generated secret masks (Compose skips that file once -f is given).
+#
+# What the base compose.yaml gives you:
+#   - `indexer` and `indexer-state-init` have no profile and no provider dependency, so they start for every
+#     provider. This fragment points the indexer at your service and makes it wait for it.
+#   - `litellm` and `hindsight` carry the profile "hindsight" and do not start under any other profile.
+# Keep this fragment small: one service (or a few), its volumes, and the indexer wiring. Never copy the indexer
+# build or image; `openmem-indexer:local` is shared.
+
+services:
+  example:
+    profiles: ["example"]
+    # Pin a tag or digest. Bump deliberately.
+    image: example/memory-server:0.0.0
+    # Loopback only. Host port from .env (EXAMPLE_PORT); change the default to the provider's usual port + 10000.
+    ports:
+      - "127.0.0.1:${EXAMPLE_PORT:-18080}:8080"
+    # Secrets live in an env file outside the repo, like HINDSIGHT_ENV_FILE. Delete if the provider needs none.
+    # required: false lets `config` and `up` run before the wizard has written the file (Compose 2.24+).
+    env_file:
+      - path: ${EXAMPLE_ENV_FILE:-${HOME}/.config/personal-memory/example.env}
+        required: false
+    volumes:
+      - example-data:/data
+    restart: unless-stopped
+    healthcheck:
+      # The image has no curl in most cases; use the interpreter that is there.
+      test: ["CMD", "python3", "-c", "import urllib.request; urllib.request.urlopen('http://127.0.0.1:8080/health', timeout=5)"]
+      interval: 15s
+      timeout: 10s
+      retries: 10
+      start_period: 30s
+
+  indexer:
+    environment:
+      # Selects indexer/sinks/example.py and gives it the API base URL (the compose service name, not localhost).
+      MEMORY_SINK: example
+      MEMORY_API_URL: http://example:8080
+    depends_on:
+      # Hard dependency (required defaults to true): this fragment is only layered together with --profile example,
+      # so a missing profile fails loudly instead of starting an indexer with nothing to write to.
+      example:
+        condition: service_healthy
+
+volumes:
+  example-data: