/ website / docs / developer-guide / memory-provider-plugin.md
memory-provider-plugin.md
  1  ---
  2  sidebar_position: 8
  3  title: "Memory Provider Plugins"
  4  description: "How to build a memory provider plugin for Hermes Agent"
  5  ---
  6  
  7  # Building a Memory Provider Plugin
  8  
  9  Memory provider plugins give Hermes Agent persistent, cross-session knowledge beyond the built-in MEMORY.md and USER.md. This guide covers how to build one.
 10  
 11  :::tip
 12  Memory providers are one of two **provider plugin** types. The other is [Context Engine Plugins](/docs/developer-guide/context-engine-plugin), which replace the built-in context compressor. Both follow the same pattern: single-select, config-driven, managed via `hermes plugins`.
 13  :::
 14  
 15  ## Directory Structure
 16  
 17  Each memory provider lives in `plugins/memory/<name>/`:
 18  
 19  ```
 20  plugins/memory/my-provider/
 21  ├── __init__.py      # MemoryProvider implementation + register() entry point
 22  ├── plugin.yaml      # Metadata (name, description, hooks)
 23  └── README.md        # Setup instructions, config reference, tools
 24  ```
 25  
 26  ## The MemoryProvider ABC
 27  
 28  Your plugin implements the `MemoryProvider` abstract base class from `agent/memory_provider.py`:
 29  
 30  ```python
 31  from agent.memory_provider import MemoryProvider
 32  
 33  class MyMemoryProvider(MemoryProvider):
 34      @property
 35      def name(self) -> str:
 36          return "my-provider"
 37  
 38      def is_available(self) -> bool:
 39          """Check if this provider can activate. NO network calls."""
 40          return bool(os.environ.get("MY_API_KEY"))
 41  
 42      def initialize(self, session_id: str, **kwargs) -> None:
 43          """Called once at agent startup.
 44  
 45          kwargs always includes:
 46            hermes_home (str): Active HERMES_HOME path. Use for storage.
 47          """
 48          self._api_key = os.environ.get("MY_API_KEY", "")
 49          self._session_id = session_id
 50  
 51      # ... implement remaining methods
 52  ```
 53  
 54  ## Required Methods
 55  
 56  ### Core Lifecycle
 57  
 58  | Method | When Called | Must Implement? |
 59  |--------|-----------|-----------------|
 60  | `name` (property) | Always | **Yes** |
 61  | `is_available()` | Agent init, before activation | **Yes** — no network calls |
 62  | `initialize(session_id, **kwargs)` | Agent startup | **Yes** |
 63  | `get_tool_schemas()` | After init, for tool injection | **Yes** |
 64  | `handle_tool_call(name, args)` | When agent uses your tools | **Yes** (if you have tools) |
 65  
 66  ### Config
 67  
 68  | Method | Purpose | Must Implement? |
 69  |--------|---------|-----------------|
 70  | `get_config_schema()` | Declare config fields for `hermes memory setup` | **Yes** |
 71  | `save_config(values, hermes_home)` | Write non-secret config to native location | **Yes** (unless env-var-only) |
 72  
 73  ### Optional Hooks
 74  
 75  | Method | When Called | Use Case |
 76  |--------|-----------|----------|
 77  | `system_prompt_block()` | System prompt assembly | Static provider info |
 78  | `prefetch(query)` | Before each API call | Return recalled context |
 79  | `queue_prefetch(query)` | After each turn | Pre-warm for next turn |
 80  | `sync_turn(user, assistant)` | After each completed turn | Persist conversation |
 81  | `on_session_end(messages)` | Conversation ends | Final extraction/flush |
 82  | `on_pre_compress(messages)` | Before context compression | Save insights before discard |
 83  | `on_memory_write(action, target, content)` | Built-in memory writes | Mirror to your backend |
 84  | `shutdown()` | Process exit | Clean up connections |
 85  
 86  ## Config Schema
 87  
 88  `get_config_schema()` returns a list of field descriptors used by `hermes memory setup`:
 89  
 90  ```python
 91  def get_config_schema(self):
 92      return [
 93          {
 94              "key": "api_key",
 95              "description": "My Provider API key",
 96              "secret": True,           # → written to .env
 97              "required": True,
 98              "env_var": "MY_API_KEY",   # explicit env var name
 99              "url": "https://my-provider.com/keys",  # where to get it
100          },
101          {
102              "key": "region",
103              "description": "Server region",
104              "default": "us-east",
105              "choices": ["us-east", "eu-west", "ap-south"],
106          },
107          {
108              "key": "project",
109              "description": "Project identifier",
110              "default": "hermes",
111          },
112      ]
113  ```
114  
115  Fields with `secret: True` and `env_var` go to `.env`. Non-secret fields are passed to `save_config()`.
116  
117  :::tip Minimal vs Full Schema
118  Every field in `get_config_schema()` is prompted during `hermes memory setup`. Providers with many options should keep the schema minimal — only include fields the user **must** configure (API key, required credentials). Document optional settings in a config file reference (e.g. `$HERMES_HOME/myprovider.json`) rather than prompting for them all during setup. This keeps the setup wizard fast while still supporting advanced configuration. See the Supermemory provider for an example — it only prompts for the API key; all other options live in `supermemory.json`.
119  :::
120  
121  ## Save Config
122  
123  ```python
124  def save_config(self, values: dict, hermes_home: str) -> None:
125      """Write non-secret config to your native location."""
126      import json
127      from pathlib import Path
128      config_path = Path(hermes_home) / "my-provider.json"
129      config_path.write_text(json.dumps(values, indent=2))
130  ```
131  
132  For env-var-only providers, leave the default no-op.
133  
134  ## Plugin Entry Point
135  
136  ```python
137  def register(ctx) -> None:
138      """Called by the memory plugin discovery system."""
139      ctx.register_memory_provider(MyMemoryProvider())
140  ```
141  
142  ## plugin.yaml
143  
144  ```yaml
145  name: my-provider
146  version: 1.0.0
147  description: "Short description of what this provider does."
148  hooks:
149    - on_session_end    # list hooks you implement
150  ```
151  
152  ## Threading Contract
153  
154  **`sync_turn()` MUST be non-blocking.** If your backend has latency (API calls, LLM processing), run the work in a daemon thread:
155  
156  ```python
157  def sync_turn(self, user_content, assistant_content):
158      def _sync():
159          try:
160              self._api.ingest(user_content, assistant_content)
161          except Exception as e:
162              logger.warning("Sync failed: %s", e)
163  
164      if self._sync_thread and self._sync_thread.is_alive():
165          self._sync_thread.join(timeout=5.0)
166      self._sync_thread = threading.Thread(target=_sync, daemon=True)
167      self._sync_thread.start()
168  ```
169  
170  ## Profile Isolation
171  
172  All storage paths **must** use the `hermes_home` kwarg from `initialize()`, not hardcoded `~/.hermes`:
173  
174  ```python
175  # CORRECT — profile-scoped
176  from hermes_constants import get_hermes_home
177  data_dir = get_hermes_home() / "my-provider"
178  
179  # WRONG — shared across all profiles
180  data_dir = Path("~/.hermes/my-provider").expanduser()
181  ```
182  
183  ## Testing
184  
185  See `tests/agent/test_memory_plugin_e2e.py` for the complete E2E testing pattern using a real SQLite provider.
186  
187  ```python
188  from agent.memory_manager import MemoryManager
189  
190  mgr = MemoryManager()
191  mgr.add_provider(my_provider)
192  mgr.initialize_all(session_id="test-1", platform="cli")
193  
194  # Test tool routing
195  result = mgr.handle_tool_call("my_tool", {"action": "add", "content": "test"})
196  
197  # Test lifecycle
198  mgr.sync_all("user msg", "assistant msg")
199  mgr.on_session_end([])
200  mgr.shutdown_all()
201  ```
202  
203  ## Adding CLI Commands
204  
205  Memory provider plugins can register their own CLI subcommand tree (e.g. `hermes my-provider status`, `hermes my-provider config`). This uses a convention-based discovery system — no changes to core files needed.
206  
207  ### How it works
208  
209  1. Add a `cli.py` file to your plugin directory
210  2. Define a `register_cli(subparser)` function that builds the argparse tree
211  3. The memory plugin system discovers it at startup via `discover_plugin_cli_commands()`
212  4. Your commands appear under `hermes <provider-name> <subcommand>`
213  
214  **Active-provider gating:** Your CLI commands only appear when your provider is the active `memory.provider` in config. If a user hasn't configured your provider, your commands won't show in `hermes --help`.
215  
216  ### Example
217  
218  ```python
219  # plugins/memory/my-provider/cli.py
220  
221  def my_command(args):
222      """Handler dispatched by argparse."""
223      sub = getattr(args, "my_command", None)
224      if sub == "status":
225          print("Provider is active and connected.")
226      elif sub == "config":
227          print("Showing config...")
228      else:
229          print("Usage: hermes my-provider <status|config>")
230  
231  def register_cli(subparser) -> None:
232      """Build the hermes my-provider argparse tree.
233  
234      Called by discover_plugin_cli_commands() at argparse setup time.
235      """
236      subs = subparser.add_subparsers(dest="my_command")
237      subs.add_parser("status", help="Show provider status")
238      subs.add_parser("config", help="Show provider config")
239      subparser.set_defaults(func=my_command)
240  ```
241  
242  ### Reference implementation
243  
244  See `plugins/memory/honcho/cli.py` for a full example with 13 subcommands, cross-profile management (`--target-profile`), and config read/write.
245  
246  ### Directory structure with CLI
247  
248  ```
249  plugins/memory/my-provider/
250  ├── __init__.py      # MemoryProvider implementation + register()
251  ├── plugin.yaml      # Metadata
252  ├── cli.py           # register_cli(subparser) — CLI commands
253  └── README.md        # Setup instructions
254  ```
255  
256  ## Single Provider Rule
257  
258  Only **one** external memory provider can be active at a time. If a user tries to register a second, the MemoryManager rejects it with a warning. This prevents tool schema bloat and conflicting backends.