--- sidebar_position: 8 title: "Memory Provider Plugins" description: "How to build a memory provider plugin for Hermes Agent" --- # Building a Memory Provider Plugin 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. :::tip Memory providers are one of two **provider plugin** types. The other is [Context Engine Plugins](./context-engine-plugin.md), which replace the built-in context compressor. Both follow the same pattern: single-select, config-driven, managed via `hermes plugins`. ::: ## Installation Layouts Hermes discovers memory providers from four sources, in this precedence order: | Source | Location | Notes | |---|---|---| | Bundled | `plugins/memory//` | Ships with Hermes. Closed to new providers — see [CONTRIBUTING](https://github.com/NousResearch/hermes-agent/blob/main/CONTRIBUTING.md). | | User | `$HERMES_HOME/plugins//` | Dropped in by the user, per profile. | | Project | `./.hermes/plugins//` | Opt-in via `HERMES_ENABLE_PROJECT_PLUGINS=1`. | | Package | `hermes_agent.memory_providers` entry point | Distribution supplied by the installation owner; nothing to copy. | Earlier sources win on a name collision, so a directory dropped into a working tree can never shadow a shipped provider. :::note This is the reverse of the general plugin system's later-wins order. A memory provider is activated by *name* (`memory.provider`), so shadowing would silently redirect the agent's memory rather than merely override a tool. ::: Discovery only *enumerates* — it never imports a provider. Nothing runs until `memory.provider` names it. Entry-point discovery does not install packages. Do not inject a provider into Hermes's selected environment with pip. On PM-managed installations, ship a directory provider with declared Python dependencies; plugin admission and `hermes memory setup` prepare them through PM before use. Owner-managed builds (such as Nix) can include an entry-point distribution declaratively. CLI and dashboard setup share candidate preparation. PM includes the provider's `pyproject.toml` or legacy `pip_dependencies` / `python_dependencies` alongside the active plugin union; an importable module does not bypass declared version constraints. Dashboard readiness checks the same inputs without installing anything. A successful preparation may require restarting Hermes before the running process can use the selected dependency generation. External sidecar checks and setup commands remain separate from the Python union. ### Directory Provider A directory provider lives in `plugins/memory//` when bundled with Hermes, in `$HERMES_HOME/plugins//` when installed by a user, or in `./.hermes/plugins//` for a project-local one: ``` plugins/memory/my-provider/ ├── __init__.py # MemoryProvider implementation + register() entry point ├── plugin.yaml # Metadata (name, description, hooks) └── README.md # Setup instructions, config reference, tools ``` ### Packaged Provider A pip-installed provider publishes an entry point in the `hermes_agent.memory_providers` group. The entry-point name is the provider name users select in `memory.provider`; its value points to the provider's `register(ctx)` function: ```toml title="pyproject.toml" [project.entry-points."hermes_agent.memory_providers"] my-provider = "my_provider:register" ``` Point the entry point at the **package**, or at a `register(ctx)` inside it, and keep your implementation, skills, and other resources in the normal Python package layout. No copy under `$HERMES_HOME/plugins/` is required. A package entry point gets everything a directory install does, including the two files Hermes reads from disk rather than importing — `config_schema.py` (the dashboard config panel) and `cli.py` (your `hermes ` subcommands). Both are found next to your package's `__init__.py`, so point the entry point at a package rather than a single module if you ship either. ## The MemoryProvider ABC Your plugin implements the `MemoryProvider` abstract base class from `agent/memory_provider.py`: ```python from agent.memory_provider import MemoryProvider from agent.secret_scope import get_secret class MyMemoryProvider(MemoryProvider): @property def name(self) -> str: return "my-provider" def is_available(self) -> bool: """Check if this provider can activate. NO network calls.""" return bool(get_secret("MY_API_KEY")) def initialize(self, session_id: str, **kwargs) -> None: """Called once at agent startup. kwargs always includes: hermes_home (str): Active HERMES_HOME path. Use for storage. """ self._api_key = get_secret("MY_API_KEY") or "" self._session_id = session_id # ... implement remaining methods ``` ### Initialization context `AIAgent` passes session context through `MemoryManager.initialize_all()` to `initialize(session_id, **kwargs)`. Accept `**kwargs` and tolerate missing optional fields; callers may initialize a provider without an agent or a session database. | Keyword | Meaning | |---|---| | `hermes_home` | Active profile's storage directory. | | `platform` | Session surface, such as `cli`, `gui`, `acp`, or `telegram`. | | `session_title` | Stored session title, when available. A display label is not necessarily a user-selected identity. | | `session_title_source` | Stored title provenance, when available: `derived`, `llm`, or `user`. Automatic sources must not be mistaken for explicit identity overrides; missing provenance retains a provider's legacy behavior. Shared constants live in `hermes_state_common.py`. | | `cwd` | Non-empty logical workspace supplied as `AIAgent(cwd=...)`, available before provider initialization. Omitted for `None` or an empty string. | | `gateway_session_key` | Stable messaging-chat identity for per-chat session isolation. | | `user_id`, `user_id_alt`, `user_name`, `chat_id` | Gateway identity fields, included when present. | | `agent_identity` | Active profile name, when available. | | `agent_workspace`, `agent_context` | Runtime agent scope. `agent_workspace` is `hermes`; `agent_context` is `cron` for scheduler runs, `subagent` for `delegate_task` children, else `primary` — skip automatic writes for the non-primary values. | Do not assume `os.getcwd()` identifies the conversation's workspace: one Desktop or gateway backend can serve several sessions. If `cwd` is absent and directory routing is needed, `agent.runtime_cwd.resolve_agent_cwd()` honors the session cwd context, then scoped `terminal.cwd` (carried internally as `TERMINAL_CWD`), then the launch directory. Construction-time workspace metadata does not require changing the process cwd or rebuilding an existing conversation's system prompt. Desktop/TUI workspace changes synchronize the live agent's `session_cwd` through `tui_gateway/session_workdir.py::_register_session_cwd`, including when a deferred agent is attached after a workspace move. This lets a first or restarted Codex app-server session use the current workspace instead of its construction-time cwd. It does not move an already-running Codex thread, reinitialize memory providers, change an existing Honcho session identity, or invalidate the cached system prompt. Provider initialization still receives the construction-time workspace; absent or empty cwd remains unpinned. ## Required Methods ### Core Lifecycle | Method | When Called | Must Implement? | |--------|-----------|-----------------| | `name` (property) | Always | **Yes** | | `is_available()` | Agent init, before activation | **Yes** — no network calls | | `initialize(session_id, **kwargs)` | Agent startup | **Yes** | | `get_tool_schemas()` | After init, for tool injection | **Yes** | | `handle_tool_call(tool_name, args, **kwargs)` | When agent uses your tools | **Yes** (if you have tools) | ### Config | Method | Purpose | Must Implement? | |--------|---------|-----------------| | `get_config_schema()` | Declare config fields for `hermes memory setup` | **Yes** | | `save_config(values, hermes_home)` | Write non-secret config to native location | **Yes** (unless env-var-only) | ### Optional Hooks | Method | When Called | Use Case | |--------|-----------|----------| | `system_prompt_block()` | System prompt assembly | Static provider info | | `prefetch(query, *, session_id="")` | Before each API call | Return recalled context | | `queue_prefetch(query, *, session_id="")` | After each turn | Pre-warm for next turn | | `sync_turn(user, assistant, *, session_id="", messages=None)` | After each completed turn | Persist conversation | | `on_session_end(messages)` | Conversation ends | Final extraction/flush | | `on_pre_compress(messages)` | Before context compression | Save insights before discard | | `on_memory_write(action, target, content, metadata=None)` | Built-in memory writes | Mirror to your backend | | `shutdown()` | Process exit | Clean up connections | For native `replace` and `remove`, `metadata["previous_content"]` contains the full entry selected under the native-store lock. Notifications are emitted only after the complete write or batch succeeds. Batch notifications preserve operation order; each operation's previous content reflects earlier operations in that batch. `old_text` is the caller's search text, not the identity of the changed entry. Older Hermes versions can omit `previous_content`. Providers that require exact identity should skip destructive mirroring when it is absent. ### Oversized prefetch results External `prefetch()` results are returned in full by default, preserving the provider's relevance-ranked recall, up to a safety ceiling of 10× `hooks.output_spill.max_chars` (at least 100,000 characters) above which they still spill. Keep your own recall budget well under that. To opt into spilling oversized results for the active profile, set: ```yaml memory: prefetch_spill_enabled: true # default: false ``` When enabled, results above the shared `hooks.output_spill.max_chars` threshold (default `10,000` characters) are written to a private spill file and replaced with the configured head/tail preview plus its path. The default preview keeps only the first and last `500` characters; the model must read the file to recover the middle. Results at or below the threshold remain unchanged. The shared `hooks.output_spill` preview lengths and directory still apply, and `hooks.output_spill.enabled: false` disables spilling even with memory opt-in. Both settings are snapshotted when the external provider is registered; restart Hermes to apply changes to existing sessions. Built-in memory and normal plugin-hook spilling are unaffected. See [Plugins — oversized-context spill](./plugins/index.md#oversized-context-spill). ## Pre-Compress Checkpoints (fail-closed) `on_pre_compress()` is best-effort by default: if your provider raises, the host logs the failure and compression proceeds. That is the right default for insight extraction — and the wrong one for a provider whose job is to archive transcript evidence to a durable store *before* the lossy rewrite. For that case the host offers an opt-in checkpoint contract (API v2): ```python from agent.memory_provider import MemoryProvider class MyArchivingProvider(MemoryProvider): # Opt in: every successful on_pre_compress() return means the durable # checkpoint is committed. Raise on any failure — do not return partial # success. Version 1 (the inherited default) is the implicit historical # contract: best-effort semantics, raw message list. pre_compress_checkpoint_api_version = 2 def on_pre_compress(self, messages, *, require_checkpoint=False): # require_checkpoint mirrors the operator's checkpoint_required # setting: True means a raise here blocks the lossy rewrite. ids = self._archive(messages) # must be durable before returning return f"checkpoint: {ids}" # forwarded into the summary prompt ``` Operators enable enforcement per deployment: ```yaml compression: checkpoint_required: true # default: false ``` With the gate on, compression **fails closed** before any lossy rewrite unless an active provider advertising the API completed its checkpoint: the uncompressed transcript is preserved, the compaction attempt errors with `BLOCKED_MISSING_PREREQUISITE`, and it can be retried once your store recovers. With the gate off (default), nothing changes for existing providers. None of the providers bundled with Hermes advertise checkpoint API v2 — the contract is opt-in and exists for third-party archiving providers. Enabling `checkpoint_required` without one therefore blocks every compression attempt (manual and automatic): agent init logs a warning naming the active provider, and each refusal names `compression.checkpoint_required` as the key to disable. The gate binds to every compaction authority, not just the Hermes summarizer: server-side native compaction (`compression.codex_responses_native`) is suppressed while the gate is armed, post-turn micro-compaction (`compression.micro_compact`) is forced off at agent init (it absorbs old exchanges into a rolling summary with no checkpoint hook in its path), and the `codex_app_server` API mode is refused at agent init — the codex agent compacts its own thread with no truthful pre-compaction boundary, so a required checkpoint cannot be guaranteed there. The checkpoint-aware Hermes compressor stays the only lossy authority. What your provider receives depends on its declared API version. Version 1 providers (the implicit default — every pre-existing provider) keep the historical contract: the raw message list, exactly as before. Version 2 checkpoint providers receive normalized direct evidence instead: user/assistant text rows only — tool results, system messages, the `tool_calls` payload of assistant messages (their prose is kept), and prior compaction summaries are filtered host-side. Prior summaries are recognized via a persistent `_compressed_summary` message marker that survives process restarts, so a resumed session never feeds derivative summaries back into your archive. **Checkpoints must be idempotent.** After a fail-closed block, the next compaction attempt calls `on_pre_compress()` again with the same transcript — and a transcript that grew only slightly produces largely overlapping evidence. Key your archive writes by content (for example a transcript digest) and upsert, so retries and overlaps deduplicate instead of accumulating duplicate archives. Contract tests: `tests/agent/test_pre_compress_checkpoint_contract.py`. ## Setup UX — what a standalone provider keeps Every setup surface Hermes gives a bundled provider is driven by files in the provider's own directory, so a provider installed from the plugin catalog keeps all of them: | Surface | What the provider ships | |---|---| | Desktop → Capabilities → Tools → Memory (config panel) | `config_schema.py` (below) | | `hermes memory setup` wizard | `get_config_schema()` declares the fields the wizard prompts for, `save_config(config, hermes_home)` persists them, `post_setup(hermes_home, config)` runs afterwards for anything interactive (OAuth, first sync); `get_status_config()` feeds `hermes memory status` | | `hermes …` subcommands | `cli.py` with `register_cli(subparser)` ([Adding CLI Commands](#adding-cli-commands)) | | Python dependencies | `pyproject.toml` `[project] dependencies` (or `python_dependencies` in `plugin.yaml`); installed under Hermes' own pins at install time and re-applied across `hermes update` | Your provider's name, `memory.` config section, data directory and tool names are the contract with existing users. A provider that moves out of core keeps all four; Hermes then installs the catalog plugin automatically for anyone whose `memory.provider` still names it (on `hermes update`, and once at agent start when `security.allow_lazy_installs` is on). ## Config Schema `get_config_schema()` returns a list of field descriptors used by `hermes memory setup`: ```python def get_config_schema(self): return [ { "key": "api_key", "description": "My Provider API key", "secret": True, # → written to .env "required": True, "env_var": "MY_API_KEY", # explicit env var name "url": "https://my-provider.com/keys", # where to get it }, { "key": "region", "description": "Server region", "default": "us-east", "choices": ["us-east", "eu-west", "ap-south"], }, { "key": "project", "description": "Project identifier", "default": "hermes", }, ] ``` Fields with `secret: True` and `env_var` go to `.env`. Non-secret fields are passed to `save_config()`. :::tip Minimal vs Full Schema 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](https://github.com/supermemoryai/hermes-supermemory) (a plugin catalog entry) for an example — it only prompts for the API key; all other options live in `supermemory.json`. ::: ## Save Config ```python def save_config(self, values: dict, hermes_home: str) -> None: """Write non-secret config to your native location.""" import json from pathlib import Path config_path = Path(hermes_home) / "my-provider.json" config_path.write_text(json.dumps(values, indent=2)) ``` For env-var-only providers, leave the default no-op. ## Plugin Entry Point ```python def register(ctx) -> None: """Called by the memory plugin discovery system.""" ctx.register_memory_provider(MyMemoryProvider()) ``` A provider may also expose read-only skills from the same callback. Skills are qualified by the entry-point name and are loaded only when that memory provider is active: ```python from pathlib import Path SKILLS_DIR = Path(__file__).parent / "skills" def register(ctx) -> None: ctx.register_memory_provider(MyMemoryProvider()) ctx.register_skill( "maintenance", SKILLS_DIR / "maintenance" / "SKILL.md", "Maintain the provider's memory store", ) ``` With the `my-provider` entry point active, the skill is available as `my-provider:maintenance` through `skill_view()`. ## plugin.yaml ```yaml name: my-provider version: 1.0.0 description: "Short description of what this provider does." hooks: - on_session_end # list hooks you implement ``` ## Threading Contract **`sync_turn()` MUST be non-blocking.** If your backend has latency (API calls, LLM processing), run the work in a daemon thread — spawned with `agent.memory_provider.spawn_context_thread`, never a bare `threading.Thread`. Profile isolation (the active `HERMES_HOME`, the per-turn secret scope) lives in `contextvars`, and a plain thread starts with an empty context: under multiplexed profiles it would silently write into the *default* profile's store, and `get_secret()` fails closed there. ```python from agent.memory_provider import spawn_context_thread def sync_turn(self, user_content, assistant_content, *, session_id="", messages=None): def _sync(): try: self._api.ingest(user_content, assistant_content, session_id=session_id, messages=messages) except Exception as e: logger.warning("Sync failed: %s", e) if self._sync_thread and self._sync_thread.is_alive(): self._sync_thread.join(timeout=5.0) self._sync_thread = spawn_context_thread(_sync, name="myprovider-sync") self._sync_thread.start() ``` The same applies to prefetch and writer threads. Small JSON config sidecars (`$HERMES_HOME/.json`) are read with `utils.read_json_or_empty` and written with `utils.atomic_json_write`; anything under `config.yaml` goes through `hermes_cli.config.save_config(..., merge_existing=True)`. `messages` is optional OpenAI-style conversation context as of the completed turn. When present, it includes user/assistant messages, assistant tool calls, and tool result messages. Providers that do not need raw turn context can omit the `messages` parameter; Hermes will continue calling them with the legacy signature. Cloud providers should document what parts of `messages` are sent off-device. Tool calls and tool results may contain file paths, command output, or other workspace data. ## Profile Isolation All storage paths **must** use the `hermes_home` kwarg from `initialize()`, not hardcoded `~/.hermes`: ```python # CORRECT — profile-scoped, bound on the instance def initialize(self, session_id, **kwargs): self._data_dir = Path(kwargs["hermes_home"]) / "my-provider" # WRONG — shared across all profiles data_dir = Path("~/.hermes/my-provider").expanduser() ``` ### Multiplexed gateways One gateway process can serve several profiles at once, and their turns run concurrently on different threads. Each agent gets its own `MemoryManager` and its own provider instance, initialized for one profile. Your provider must: - **Return a new instance per load.** `register(ctx)` should construct the provider (`ctx.register_memory_provider(MyProvider())`), not hand back a module-level singleton. The manager refuses to initialize an instance that another profile's manager already bound, and logs `memory_provider_instance_shared`. - **Bind from `initialize()`.** Derive every store, path, client, bank and tenant from the `hermes_home` keyword (and secrets read through `get_secret()` during `initialize`), and keep that binding on `self`. - **Keep nothing profile-specific at module or class level.** No global "current store", and no rebinding a shared binding on each call. Two profiles' calls interleave, so a per-call rebind races. The manager enforces the other half. It runs tool calls and every lifecycle hook (`prefetch`, `queue_prefetch`, `sync_turn`, `on_memory_write`, `on_session_end`, …) only from the profile the instance was initialized for. On a multiplexed gateway a caller that names no profile is refused, and the refusal is logged as `memory_caller_profile_unresolved` in the profile's `errors.log` (`hermes logs errors`) and in `gateway.log`. Background hooks run under the owning profile's `HERMES_HOME`. `handle_tool_call(..., **kwargs)` also receives `hermes_home`, `agent_identity` and `caller_gsk`. They name the profile your instance is already bound to, so you can assert on them, but don't rebind from them. Hermes can't verify a provider's internals. A provider that keeps a process-wide binding still mixes profiles' memories, even with these checks in place. ## Testing See `tests/agent/test_memory_provider.py` and adjacent memory tests (`tests/agent/test_memory_session_switch.py`, `tests/agent/test_memory_user_id.py`, `tests/agent/test_memory_provider_init.py`) for end-to-end patterns. ```python from agent.memory_manager import MemoryManager mgr = MemoryManager() mgr.add_provider(my_provider) mgr.initialize_all(session_id="test-1", platform="cli") # Test tool routing result = mgr.handle_tool_call("my_tool", {"action": "add", "content": "test"}) # Test lifecycle mgr.sync_all("user msg", "assistant msg") mgr.on_session_end([]) mgr.shutdown_all() ``` ## Adding CLI Commands 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. ### How it works 1. Add a `cli.py` file to your plugin directory 2. Define a `register_cli(subparser)` function that builds the argparse tree 3. The memory plugin system discovers it at startup via `discover_plugin_cli_commands()` 4. Your commands appear under `hermes ` **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`. ### Example ```python # plugins/memory/my-provider/cli.py def my_command(args): """Handler dispatched by argparse.""" sub = getattr(args, "my_command", None) if sub == "status": print("Provider is active and connected.") elif sub == "config": print("Showing config...") else: print("Usage: hermes my-provider ") def register_cli(subparser) -> None: """Build the hermes my-provider argparse tree. Called by discover_plugin_cli_commands() at argparse setup time. """ subs = subparser.add_subparsers(dest="my_command") subs.add_parser("status", help="Show provider status") subs.add_parser("config", help="Show provider config") subparser.set_defaults(func=my_command) ``` ### Reference implementation See the Honcho plugin's [`cli.py`](https://github.com/plastic-labs/honcho/blob/main/hermes-plugin-honcho/cli.py) for a full example with 13 subcommands, cross-profile management (`--target-profile`), and config read/write. ### Directory structure with CLI ``` plugins/memory/my-provider/ ├── __init__.py # MemoryProvider implementation + register() ├── plugin.yaml # Metadata ├── cli.py # register_cli(subparser) — CLI commands └── README.md # Setup instructions ``` ## Single Provider Rule 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. ## `HERMES_HOME` survival contract (what wrappers can rely on) For wrapper-style providers that keep their runtime in a sidecar venv outside Hermes-managed Python (no dependency surface — no `pyproject.toml`, `pip_dependencies`, or `python_dependencies` — at the scanned plugin root; a `pyproject.toml` belonging solely to an external or nested sidecar is not scanned): - **Location.** `$HERMES_HOME/plugins//` is the profile-scoped plugin location, and `HERMES_HOME` follows the active context override, then `$HERMES_HOME`, then the platform default. Propagate `HERMES_HOME` when launching the wrapper or sidecar so profile isolation holds; `MemoryManager.initialize_all` injects the active `hermes_home` into every provider. - **Survival.** Ordinary Hermes updates — including managed-venv rebuild/replacement by pm — do not delete or rewrite `$HERMES_HOME/plugins/**`. An installed wrapper directory and its marker file (e.g. `mnemosyne-wrapper.json`) survive. Explicit plugin updates and deletion flows (`hermes uninstall`, `hermes plugins remove`, profile deletion, user deletion) are excluded from this guarantee. - **Sidecar isolation.** A plugin root with no dependency surface never joins the pm workspace dependency union; a resync or venv rebuild neither provisions deps for it nor touches its tree. - **Conflicts.** For native shared-venv plugins, an unsatisfiable dependency union fails loudly: the candidate plugin stays unenabled and unimported (the admission authority refuses before publishing config, reporting the plugin identity plus the resolver's reason, with a re-enable/retry path and a machine-readable pm receipt). Dependency resolution does not automatically disable other plugins or run a bisect. Explicit plugin updates, removal, and independent security gates are separate operations.