diff --git a/reme4/components/base_component.py b/reme4/components/base_component.py
index 06e289e3..4c614d52 100644
--- a/reme4/components/base_component.py
+++ b/reme4/components/base_component.py
@@ -142,6 +142,12 @@ class BaseComponent(ABC):
     async def _close(self) -> None:
         """Subclass hook: close logic."""
 
+    async def dump(self) -> None:
+        """Persist in-memory state to disk. Override in subclasses that need persistence."""
+
+    async def load(self) -> None:
+        """Restore in-memory state from disk. Override in subclasses that need persistence."""
+
     async def start(self) -> None:
         """Resolve bindings → start owned fallbacks → _start(). No-op if already started."""
         async with self._lock:
diff --git a/reme4/components/embedding/base_embedding_model.py b/reme4/components/embedding/base_embedding_model.py
index 4fce0535..02e2c877 100644
--- a/reme4/components/embedding/base_embedding_model.py
+++ b/reme4/components/embedding/base_embedding_model.py
@@ -56,8 +56,7 @@ class BaseEmbeddingModel(BaseComponent):
 
     async def _start(self) -> None:
         """Load cache from disk on startup."""
-        self._embedding_cache.clear()
-        self._load_cache()
+        await self.load()
 
     async def health_check(self, timeout: float = 2.0) -> bool:
         """Probe the provider; sets and returns is_healthy."""
@@ -78,7 +77,7 @@ class BaseEmbeddingModel(BaseComponent):
 
     async def _close(self) -> None:
         """Persist cache to disk on shutdown."""
-        self._save_cache()
+        await self.dump()
 
     # -- Public API --
 
@@ -180,12 +179,10 @@ class BaseEmbeddingModel(BaseComponent):
 
     # -- Cache Persistence --
 
-    def _load_cache(self) -> None:
-        """Load cached embeddings from disk (npz format)."""
-        if not self.enable_cache:
-            return
-        self.cache_path.parent.mkdir(parents=True, exist_ok=True)
-        if not self.cache_path.exists():
+    async def load(self) -> None:
+        """Load cached embeddings from disk (npz format); replaces in-memory cache."""
+        self._embedding_cache.clear()
+        if not self.enable_cache or not self.cache_path.exists():
             return
 
         try:
@@ -202,10 +199,11 @@ class BaseEmbeddingModel(BaseComponent):
                 break
             self._embedding_cache[str(key)] = emb.astype(np.float16)
 
-    def _save_cache(self) -> None:
+    async def dump(self) -> None:
         """Persist in-memory cache to disk (npz format)."""
         if not self.enable_cache or not self._embedding_cache:
             return
+        self.cache_path.parent.mkdir(parents=True, exist_ok=True)
         keys = list(self._embedding_cache.keys())
         embeddings = np.stack(list(self._embedding_cache.values()))
         try:
diff --git a/reme4/components/file_graph/local_file_graph.py b/reme4/components/file_graph/local_file_graph.py
index 8aed20d7..221b022a 100644
--- a/reme4/components/file_graph/local_file_graph.py
+++ b/reme4/components/file_graph/local_file_graph.py
@@ -22,7 +22,7 @@ class LocalFileGraph(BaseFileGraph):
 
     async def _start(self) -> None:
         await super()._start()
-        self._load()
+        await self.load()
         await self.rebuild_links()
         self.logger.info(
             f"LocalFileGraph '{self.graph_name}' ready: "
@@ -31,22 +31,30 @@ class LocalFileGraph(BaseFileGraph):
         )
 
     async def _close(self) -> None:
-        self._dump()
+        await self.dump()
         await super()._close()
 
-    def _load(self) -> None:
-        """Load nodes from JSONL file into memory."""
+    async def load(self) -> None:
+        """Load nodes from JSONL file into memory; keep current state on failure."""
         if not self._graph_file.exists():
             return
-        with open(self._graph_file, "r", encoding="utf-8") as f:
-            self._nodes.update((n.path, n) for line in f if line.strip() for n in [FileNode.model_validate_json(line)])
+        try:
+            with open(self._graph_file, "r", encoding="utf-8") as f:
+                self._nodes.update(
+                    (n.path, n) for line in f if line.strip() for n in [FileNode.model_validate_json(line)]
+                )
+        except Exception as e:
+            self.logger.exception(f"Failed to load {self._graph_file}: {e}")
 
-    def _dump(self) -> None:
+    async def dump(self) -> None:
         """Persist all nodes to JSONL via atomic rename."""
-        tmp = self._graph_file.with_suffix(".tmp")
-        with open(tmp, "w", encoding="utf-8") as f:
-            f.writelines(f"{n.model_dump_json()}\n" for n in self._nodes.values())
-        tmp.replace(self._graph_file)
+        try:
+            tmp = self._graph_file.with_suffix(".tmp")
+            with open(tmp, "w", encoding="utf-8") as f:
+                f.writelines(f"{n.model_dump_json()}\n" for n in self._nodes.values())
+            tmp.replace(self._graph_file)
+        except Exception as e:
+            self.logger.exception(f"Failed to write {self._graph_file}: {e}")
 
     # -- Edge bookkeeping --------------------------------------------------
 
diff --git a/reme4/components/file_graph/nx_file_graph.py b/reme4/components/file_graph/nx_file_graph.py
index 81695dec..dcb608d6 100644
--- a/reme4/components/file_graph/nx_file_graph.py
+++ b/reme4/components/file_graph/nx_file_graph.py
@@ -28,9 +28,7 @@ class NxFileGraph(BaseFileGraph):
 
     async def _start(self) -> None:
         await super()._start()
-        loaded = self._load()
-        if loaded is not None:
-            self._graph = loaded
+        await self.load()
         n_real = sum(1 for _, d in self._graph.nodes(data=True) if "node" in d)
         self.logger.info(
             f"NxFileGraph '{self.graph_name}' ready: "
@@ -39,21 +37,20 @@ class NxFileGraph(BaseFileGraph):
         )
 
     async def _close(self) -> None:
-        self._dump()
+        await self.dump()
         await super()._close()
 
-    def _load(self) -> nx.MultiDiGraph | None:
-        """Load graph from pickle file; return None on failure."""
+    async def load(self) -> None:
+        """Load graph from pickle file; keep current graph on failure."""
         if not self._graph_file.exists():
-            return None
+            return
         try:
             with open(self._graph_file, "rb") as f:
-                return pickle.load(f)
+                self._graph = pickle.load(f)
         except Exception as e:
             self.logger.exception(f"Failed to load {self._graph_file}: {e}")
-            return None
 
-    def _dump(self) -> None:
+    async def dump(self) -> None:
         """Persist graph to pickle via atomic rename."""
         try:
             tmp = self._graph_file.with_suffix(".tmp")
diff --git a/reme4/components/file_store/local_file_store.py b/reme4/components/file_store/local_file_store.py
index 63638dd6..052694e4 100644
--- a/reme4/components/file_store/local_file_store.py
+++ b/reme4/components/file_store/local_file_store.py
@@ -23,19 +23,30 @@ class LocalFileStore(BaseFileStore):
 
     async def _start(self) -> None:
         await super()._start()
-        if self.chunks_path.exists():
-            try:
-                async with aiofiles.open(self.chunks_path, encoding=self.encoding) as f:
-                    async for line in f:
-                        line = line.strip()
-                        if line:
-                            chunk = FileChunk.model_validate_json(line)
-                            self.file_chunks[chunk.id] = chunk
-            except Exception as e:
-                self.logger.exception(f"Failed to load {self.chunks_path}: {e}")
+        await self.load()
         self.logger.info(f"LocalFileStore '{self.store_name}' ready: {len(self.file_chunks)} chunks")
 
     async def _close(self) -> None:
+        await self.dump()
+        self.file_chunks.clear()
+        await super()._close()
+
+    async def load(self) -> None:
+        """Load chunks from JSONL file into memory."""
+        if not self.chunks_path.exists():
+            return
+        try:
+            async with aiofiles.open(self.chunks_path, encoding=self.encoding) as f:
+                async for line in f:
+                    line = line.strip()
+                    if line:
+                        chunk = FileChunk.model_validate_json(line)
+                        self.file_chunks[chunk.id] = chunk
+        except Exception as e:
+            self.logger.exception(f"Failed to load {self.chunks_path}: {e}")
+
+    async def dump(self) -> None:
+        """Persist chunks to JSONL via atomic rename."""
         try:
             tmp = self.chunks_path.with_suffix(".tmp")
             async with aiofiles.open(tmp, "w", encoding=self.encoding) as f:
@@ -43,8 +54,6 @@ class LocalFileStore(BaseFileStore):
             tmp.replace(self.chunks_path)
         except Exception as e:
             self.logger.exception(f"Failed to write {self.chunks_path}: {e}")
-        self.file_chunks.clear()
-        await super()._close()
 
     # Base class interface
 
diff --git a/reme4/components/keyword_index/base_keyword_index.py b/reme4/components/keyword_index/base_keyword_index.py
index 45681a4d..c577e878 100644
--- a/reme4/components/keyword_index/base_keyword_index.py
+++ b/reme4/components/keyword_index/base_keyword_index.py
@@ -24,9 +24,8 @@ class BaseKeywordIndex(BaseComponent):
 
     async def _start(self) -> None:
         """Load existing index from disk if available."""
-        if self.index_file.exists():
-            await self.load()
-            self.logger.info(f"Loaded index from {self.index_path}")
+        await self.load()
+        self.logger.info(f"Loaded index from {self.index_path}")
 
     async def _close(self) -> None:
         """Save index to disk on shutdown."""
@@ -59,14 +58,6 @@ class BaseKeywordIndex(BaseComponent):
     async def retrieve(self, query: str, limit: int = 3) -> dict[str, float]:
         """Search documents. Returns {doc_id: score} sorted descending."""
 
-    @abstractmethod
-    async def dump(self) -> None:
-        """Persist index to disk."""
-
-    @abstractmethod
-    async def load(self) -> None:
-        """Load index from disk."""
-
     @abstractmethod
     async def clear(self) -> None:
         """Reset index to empty state."""
diff --git a/reme4/components/keyword_index/bm25_index.py b/reme4/components/keyword_index/bm25_index.py
index c64c70f3..3c2591ac 100644
--- a/reme4/components/keyword_index/bm25_index.py
+++ b/reme4/components/keyword_index/bm25_index.py
@@ -127,22 +127,29 @@ class BM25Index(BaseKeywordIndex):
         return dict(sorted(scores.items(), key=lambda x: x[1], reverse=True)[:limit]) if scores else {}
 
     async def dump(self) -> None:
-        """Persist index to disk via pickle."""
-        with open(self.index_file, "wb") as f:
-            pickle.dump(
-                {
-                    "vocab": self.vocab,
-                    "inverted_index": self.inverted_index,
-                    "doc_meta": self.doc_meta,
-                    "total_len": self.total_len,
-                    "k1": self.k1,
-                    "b": self.b,
-                },
-                f,
-            )
+        """Persist index to disk via pickle (atomic rename)."""
+        try:
+            tmp = self.index_file.with_suffix(".tmp")
+            with open(tmp, "wb") as f:
+                pickle.dump(
+                    {
+                        "vocab": self.vocab,
+                        "inverted_index": self.inverted_index,
+                        "doc_meta": self.doc_meta,
+                        "total_len": self.total_len,
+                        "k1": self.k1,
+                        "b": self.b,
+                    },
+                    f,
+                )
+            tmp.replace(self.index_file)
+        except Exception as e:
+            self.logger.exception(f"Failed to write {self.index_file}: {e}")
 
     async def load(self) -> None:
-        """Load index from disk. Clears index on failure."""
+        """Load index from disk. No-op if file missing; clears index on corruption."""
+        if not self.index_file.exists():
+            return
         try:
             with open(self.index_file, "rb") as f:
                 data = pickle.load(f)
diff --git a/reme4/config/default.yaml b/reme4/config/default.yaml
index f06d6c5a..8bcfd5c0 100644
--- a/reme4/config/default.yaml
+++ b/reme4/config/default.yaml
@@ -40,6 +40,15 @@ jobs:
     steps:
       - backend: health_check_step
 
+  - backend: base
+    name: help
+    description: "list all registered jobs with their metadata"
+    parameters:
+      type: object
+      properties: {}
+    steps:
+      - backend: help_step
+
   - backend: stream
     name: stream_demo
     description: "stream demo job: repeat query 10x and stream char-by-char"
diff --git a/reme4/steps/common/__init__.py b/reme4/steps/common/__init__.py
index fb67df6a..e31ba302 100644
--- a/reme4/steps/common/__init__.py
+++ b/reme4/steps/common/__init__.py
@@ -2,6 +2,7 @@
 
 from .demo import DemoEchoStep1, DemoEchoStep2
 from .health_check import HealthCheckStep
+from .help import HelpStep
 from .stream_demo import StreamDemoStep1, StreamDemoStep2
 from .version import VersionStep
 
@@ -9,6 +10,7 @@ __all__ = [
     "DemoEchoStep1",
     "DemoEchoStep2",
     "HealthCheckStep",
+    "HelpStep",
     "StreamDemoStep1",
     "StreamDemoStep2",
     "VersionStep",
diff --git a/reme4/steps/common/help.py b/reme4/steps/common/help.py
new file mode 100644
index 00000000..782ea877
--- /dev/null
+++ b/reme4/steps/common/help.py
@@ -0,0 +1,42 @@
+"""Return a one-line summary of every registered job for LLM consumption."""
+
+from ..base_step import BaseStep
+from ...components import R
+
+
+def _format_params(parameters: dict) -> str:
+    props = (parameters or {}).get("properties") or {}
+    if not props:
+        return "no args"
+    required = set((parameters or {}).get("required") or [])
+    parts = []
+    for pname, pschema in props.items():
+        ptype = pschema.get("type", "any")
+        if pname in required:
+            parts.append(f"{pname}:{ptype}*")
+        elif "default" in pschema:
+            parts.append(f"{pname}:{ptype}={pschema['default']}")
+        else:
+            parts.append(f"{pname}:{ptype}")
+    return ", ".join(parts)
+
+
+@R.register("help_step")
+class HelpStep(BaseStep):
+    """List all registered jobs (excluding self) as compact one-liners for an LLM."""
+
+    async def execute(self):
+        assert self.context is not None
+
+        lines = []
+        if self.app_context is not None:
+            for name, job in self.app_context.jobs.items():
+                if name == "help":
+                    continue
+                lines.append(f"🛠️ `{name}` — {job.description} 📥 {_format_params(job.parameters)}")
+
+        self.logger.info(f"[{self.name}] returning {len(lines)} jobs")
+
+        self.context.response.answer = "\n".join(lines)
+        self.context.response.metadata["job_count"] = len(lines)
+        return self.context.response