feat: add profile-scoped local Hindsight eval path

Closes #1010

agent/context_snapshots.py

@@ -0,0 +1,69 @@
+"""First-class context snapshot artifacts for live runtime memory evaluation."""
+
+from __future__ import annotations
+
+import json
+import re
+from pathlib import Path
+from typing import Any
+
+from hermes_constants import get_hermes_home
+
+
+_SAFE_SEGMENT_RE = re.compile(r"[^A-Za-z0-9_.-]+")
+
+
+class ContextSnapshotRecorder:
+    """Write per-call prompt-composition artifacts for a Hermes session."""
+
+    def __init__(self, session_id: str, *, enabled: bool = False, base_dir: str | Path | None = None):
+        self.session_id = session_id or "session"
+        self.enabled = bool(enabled)
+        self.base_dir = Path(base_dir) if base_dir else get_hermes_home() / "reports" / "context_snapshots"
+
+    @property
+    def session_dir(self) -> Path:
+        safe_session = _SAFE_SEGMENT_RE.sub("_", self.session_id).strip("._") or "session"
+        return self.base_dir / safe_session
+
+    def record_call(
+        self,
+        api_call_count: int,
+        *,
+        system_prompt: str,
+        memory_provider_system_prompt: str = "",
+        memory_prefetch_raw: str = "",
+        memory_context_block: str = "",
+        api_user_message: str = "",
+        api_messages: list[dict[str, Any]] | None = None,
+        metadata: dict[str, Any] | None = None,
+    ) -> Path | None:
+        if not self.enabled:
+            return None
+
+        call_dir = self.session_dir / f"call_{api_call_count:03d}"
+        call_dir.mkdir(parents=True, exist_ok=True)
+
+        self._write_text(call_dir / "system_prompt.txt", system_prompt or "")
+        self._write_text(call_dir / "memory_provider_system_prompt.txt", memory_provider_system_prompt or "")
+        self._write_text(call_dir / "memory_prefetch_raw.txt", memory_prefetch_raw or "")
+        self._write_text(call_dir / "memory_context_block.txt", memory_context_block or "")
+        self._write_text(call_dir / "api_user_message.txt", api_user_message or "")
+        self._write_json(call_dir / "api_messages.json", api_messages or [])
+        self._write_json(
+            call_dir / "metadata.json",
+            {
+                "session_id": self.session_id,
+                "api_call_count": api_call_count,
+                **(metadata or {}),
+            },
+        )
+        return call_dir
+
+    @staticmethod
+    def _write_text(path: Path, content: str) -> None:
+        path.write_text(content, encoding="utf-8")
+
+    @staticmethod
+    def _write_json(path: Path, payload: Any) -> None:
+        path.write_text(json.dumps(payload, indent=2, ensure_ascii=False), encoding="utf-8")

docs/hindsight-local-eval.md

@@ -0,0 +1,132 @@
+# Hindsight local eval homes for live Hermes runtime testing
+
+Issue: #1010
+Parent: #985
+
+This document defines a reproducible, profile-scoped evaluation layout for baseline / MemPalace / Hindsight comparisons without requiring Hindsight Cloud.
+
+## Eval home layout
+
+Use three separate `HERMES_HOME` directories so each run has isolated config, memory, sessions, and artifacts.
+
+```text
+~/.hermes/profiles/atlas-baseline/
+  config.yaml
+  .env
+  MEMORY.md
+  USER.md
+  reports/context_snapshots/
+
+~/.hermes/profiles/atlas-mempalace/
+  config.yaml
+  .env
+  MEMORY.md
+  USER.md
+  reports/context_snapshots/
+  plugins/           # if a local MemPalace plugin is installed for this eval lane
+
+~/.hermes/profiles/atlas-hindsight/
+  config.yaml
+  .env
+  MEMORY.md
+  USER.md
+  hindsight/config.json
+  reports/context_snapshots/
+```
+
+## Hindsight local config
+
+The Hindsight provider already loads config from `$HERMES_HOME/hindsight/config.json` first. For the local eval lane, prefer `local_embedded` so Hermes can bring up a local Hindsight daemon without cloud signup.
+
+Example `~/.hermes/profiles/atlas-hindsight/hindsight/config.json`:
+
+```json
+{
+  "mode": "local_embedded",
+  "memory_mode": "context",
+  "recall_prefetch_method": "recall",
+  "llm_provider": "ollama",
+  "llm_model": "gemma3:12b",
+  "api_url": "http://localhost:8888"
+}
+```
+
+Notes:
+- `local_embedded` avoids any Hindsight Cloud dependency.
+- If `profile` is omitted, Hermes now derives a stable local Hindsight profile name from the active profile identity / `HERMES_HOME` instead of collapsing all local runs into the shared legacy `hermes` profile.
+- `local_external` remains valid if you already run a local Hindsight server yourself.
+
+## Runtime switching procedure
+
+Switch by exporting `HERMES_HOME` before launching Hermes.
+
+### 1. Baseline
+
+```bash
+export HERMES_HOME="$HOME/.hermes/profiles/atlas-baseline"
+unset HERMES_CONTEXT_SNAPSHOTS
+hermes chat
+```
+
+### 2. MemPalace lane
+
+```bash
+export HERMES_HOME="$HOME/.hermes/profiles/atlas-mempalace"
+export HERMES_CONTEXT_SNAPSHOTS=1
+hermes chat
+```
+
+### 3. Hindsight lane
+
+```bash
+export HERMES_HOME="$HOME/.hermes/profiles/atlas-hindsight"
+export HERMES_CONTEXT_SNAPSHOTS=1
+hermes chat
+```
+
+## Raw artifact capture
+
+When `HERMES_CONTEXT_SNAPSHOTS=1` is enabled, Hermes writes first-class prompt-composition artifacts under the active home by default.
+
+Artifact tree:
+
+```text
+$HERMES_HOME/reports/context_snapshots/<session-id>/call_001/
+  system_prompt.txt
+  memory_provider_system_prompt.txt
+  memory_prefetch_raw.txt
+  memory_context_block.txt
+  api_user_message.txt
+  api_messages.json
+  metadata.json
+```
+
+Minimum files a benchmark should inspect:
+- `system_prompt.txt`
+- `memory_prefetch_raw.txt`
+- `memory_context_block.txt`
+- `api_user_message.txt`
+- `api_messages.json`
+
+These prove:
+- what the system prompt was
+- what the provider prefetched
+- what entered `<memory-context>`
+- what the final API user message looked like
+- what full payload reached the model
+
+## Follow-on benchmark workflow
+
+A benchmark issue can now consume this path without redoing integration work:
+1. pick one eval home (`atlas-baseline`, `atlas-mempalace`, `atlas-hindsight`)
+2. export the corresponding `HERMES_HOME`
+3. run Hermes on the same prompt set
+4. compare the snapshot artifacts in `reports/context_snapshots/`
+5. score recall quality and answer quality separately
+
+## Why this is sovereign
+
+- no hosted Hindsight Cloud dependency is required
+- the Hindsight config is profile-scoped under `hindsight/config.json`
+- the runtime artifacts stay under the active `HERMES_HOME`
+- switching between baseline / MemPalace / Hindsight is just a `HERMES_HOME` swap

hermes_cli/config.py

@@ -523,7 +523,7 @@ DEFAULT_CONFIG = {
     
     # Text-to-speech configuration
     "tts": {
-        "provider": "edge",  # "edge" (free) | "elevenlabs" (premium) | "openai" | "minimax" | "mistral" | "neutts" (local) | "kittentts" (local)
+        "provider": "edge",  # "edge" (free) | "elevenlabs" (premium) | "openai" | "minimax" | "mistral" | "neutts" (local)
         "edge": {
             "voice": "en-US-AriaNeural",
             # Popular: AriaNeural, JennyNeural, AndrewNeural, BrianNeural, SoniaNeural
@@ -547,12 +547,6 @@ DEFAULT_CONFIG = {
             "model": "neuphonic/neutts-air-q4-gguf",  # HuggingFace model repo
             "device": "cpu",  # cpu, cuda, or mps
         },
-        "kittentts": {
-            "model": "KittenML/kitten-tts-nano-0.8-int8",  # 25MB int8 default
-            "voice": "Jasper",  # Jasper, Bella, Luna, Bruno, Rosie, Hugo, Kiki, Leo
-            "speed": 1.0,
-            "clean_text": True,
-        },
     },
     
     "stt": {

hermes_cli/setup.py

@@ -443,16 +443,6 @@ def _print_setup_summary(config: dict, hermes_home):
             tool_status.append(("Text-to-Speech (NeuTTS local)", True, None))
         else:
             tool_status.append(("Text-to-Speech (NeuTTS — not installed)", False, "run 'hermes setup tts'"))
-    elif tts_provider == "kittentts":
-        try:
-            import importlib.util
-            kittentts_ok = importlib.util.find_spec("kittentts") is not None
-        except Exception:
-            kittentts_ok = False
-        if kittentts_ok:
-            tool_status.append(("Text-to-Speech (KittenTTS local)", True, None))
-        else:
-            tool_status.append(("Text-to-Speech (KittenTTS — not installed)", False, "run 'hermes setup tts'"))
     else:
         tool_status.append(("Text-to-Speech (Edge TTS)", True, None))
 
@@ -901,7 +891,6 @@ def _install_neutts_deps() -> bool:
                 return False
         else:
             print_warning("espeak-ng is required for NeuTTS. Install it manually before using NeuTTS.")
-            return False
 
     # Install neutts Python package
     print()
@@ -921,34 +910,8 @@ def _install_neutts_deps() -> bool:
         return False
 
 
-def _install_kittentts_deps() -> bool:
-    """Install KittenTTS dependencies with user approval. Returns True on success."""
-    import subprocess
-    import sys
-
-    wheel_url = (
-        "https://github.com/KittenML/KittenTTS/releases/download/"
-        "0.8.1/kittentts-0.8.1-py3-none-any.whl"
-    )
-    print()
-    print_info("Installing kittentts Python package (~25-80MB model downloaded on first use)...")
-    print()
-    try:
-        subprocess.run(
-            [sys.executable, "-m", "pip", "install", "-U", wheel_url, "soundfile", "--quiet"],
-            check=True, timeout=300,
-        )
-        print_success("kittentts installed successfully")
-        return True
-    except (subprocess.CalledProcessError, subprocess.TimeoutExpired) as e:
-        print_error(f"Failed to install kittentts: {e}")
-        print_info(f"Try manually: python -m pip install -U '{wheel_url}' soundfile")
-        return False
-
-
 def _setup_tts_provider(config: dict):
-    """Interactive TTS provider selection with install flow for local providers."""
-
+    """Interactive TTS provider selection with install flow for NeuTTS."""
     tts_config = config.get("tts", {})
     current_provider = tts_config.get("provider", "edge")
     subscription_features = get_nous_subscription_features(config)
@@ -960,7 +923,6 @@ def _setup_tts_provider(config: dict):
         "minimax": "MiniMax TTS",
         "mistral": "Mistral Voxtral TTS",
         "neutts": "NeuTTS",
-        "kittentts": "KittenTTS",
     }
     current_label = provider_labels.get(current_provider, current_provider)
 
@@ -982,10 +944,9 @@ def _setup_tts_provider(config: dict):
             "MiniMax TTS (high quality with voice cloning, needs API key)",
             "Mistral Voxtral TTS (multilingual, native Opus, needs API key)",
             "NeuTTS (local on-device, free, ~300MB model download)",
-            "KittenTTS (local on-device, free, lightweight ~25-80MB ONNX)",
         ]
     )
-    providers.extend(["edge", "elevenlabs", "openai", "minimax", "mistral", "neutts", "kittentts"])
+    providers.extend(["edge", "elevenlabs", "openai", "minimax", "mistral", "neutts"])
     choices.append(f"Keep current ({current_label})")
     keep_current_idx = len(choices) - 1
     idx = prompt_choice("Select TTS provider:", choices, keep_current_idx)
@@ -1027,28 +988,6 @@ def _setup_tts_provider(config: dict):
                 print_info("Skipping install. Set tts.provider to 'neutts' after installing manually.")
                 selected = "edge"
 
-    elif selected == "kittentts":
-        try:
-            import importlib.util
-            already_installed = importlib.util.find_spec("kittentts") is not None
-        except Exception:
-            already_installed = False
-
-        if already_installed:
-            print_success("KittenTTS is already installed")
-        else:
-            print()
-            print_info("KittenTTS is lightweight (~25-80MB, CPU-only, no API key required).")
-            print_info("Voices: Jasper, Bella, Luna, Bruno, Rosie, Hugo, Kiki, Leo")
-            print()
-            if prompt_yes_no("Install KittenTTS now?", True):
-                if not _install_kittentts_deps():
-                    print_warning("KittenTTS installation incomplete. Falling back to Edge TTS.")
-                    selected = "edge"
-            else:
-                print_info("Skipping install. Set tts.provider to 'kittentts' after installing manually.")
-                selected = "edge"
-
     elif selected == "elevenlabs":
         existing = get_env_value("ELEVENLABS_API_KEY")
         if not existing:

hermes_cli/tools_config.py

@@ -164,14 +164,6 @@ TOOL_CATEGORIES = {
                 ],
                 "tts_provider": "mistral",
             },
-            {
-                "name": "KittenTTS",
-                "badge": "local · free",
-                "tag": "Lightweight local ONNX TTS (~25MB), no API key",
-                "env_vars": [],
-                "tts_provider": "kittentts",
-                "post_setup": "kittentts",
-            },
         ],
     },
     "web": {
@@ -411,36 +403,6 @@ def _run_post_setup(post_setup_key: str):
             _print_warning("    Node.js not found. Install Camofox via Docker:")
             _print_info("      docker run -p 9377:9377 -e CAMOFOX_PORT=9377 jo-inc/camofox-browser")
 
-    elif post_setup_key == "kittentts":
-        try:
-            __import__("kittentts")
-            _print_success("    kittentts is already installed")
-            return
-        except ImportError:
-            pass
-        import subprocess
-        _print_info("    Installing kittentts (~25-80MB model, CPU-only)...")
-        wheel_url = (
-            "https://github.com/KittenML/KittenTTS/releases/download/"
-            "0.8.1/kittentts-0.8.1-py3-none-any.whl"
-        )
-        try:
-            result = subprocess.run(
-                [sys.executable, "-m", "pip", "install", "-U", wheel_url, "soundfile", "--quiet"],
-                capture_output=True, text=True, timeout=300,
-            )
-            if result.returncode == 0:
-                _print_success("    kittentts installed")
-                _print_info("    Voices: Jasper, Bella, Luna, Bruno, Rosie, Hugo, Kiki, Leo")
-                _print_info("    Models: KittenML/kitten-tts-nano-0.8-int8 (25MB), micro (41MB), mini (80MB)")
-            else:
-                _print_warning("    kittentts install failed:")
-                _print_info(f"      {result.stderr.strip()[:300]}")
-                _print_info(f"    Run manually: python -m pip install -U '{wheel_url}' soundfile")
-        except subprocess.TimeoutExpired:
-            _print_warning("    kittentts install timed out (>5min)")
-            _print_info(f"    Run manually: python -m pip install -U '{wheel_url}' soundfile")
-
     elif post_setup_key == "rl_training":
         try:
             __import__("tinker_atropos")

plugins/memory/hindsight/__init__.py

@@ -178,6 +178,25 @@ def _load_config() -> dict:
     }
 
 
+def _derive_local_profile_name(agent_identity: str = "", hermes_home: str = "") -> str:
+    """Return a stable profile name for local embedded Hindsight storage.
+
+    Prefer the active Hermes profile identity when available, otherwise fall back
+    to the basename of the active HERMES_HOME path. This prevents all local
+    Hindsight eval homes from sharing the legacy default profile name "hermes".
+    """
+    from pathlib import Path
+    import re
+
+    raw = (agent_identity or "").strip()
+    if not raw and hermes_home:
+        raw = Path(hermes_home).name.strip()
+    if not raw:
+        raw = "hermes"
+    safe = re.sub(r"[^A-Za-z0-9_.-]+", "-", raw).strip(".-_")
+    return safe or "hermes"
+
+
 # ---------------------------------------------------------------------------
 # MemoryProvider implementation
 # ---------------------------------------------------------------------------
@@ -468,6 +487,8 @@ class HindsightMemoryProvider(MemoryProvider):
 
     def initialize(self, session_id: str, **kwargs) -> None:
         self._session_id = session_id
+        hermes_home = str(kwargs.get("hermes_home") or "")
+        agent_identity = str(kwargs.get("agent_identity") or "")
 
         # Check client version and auto-upgrade if needed
         try:
@@ -500,6 +521,11 @@ class HindsightMemoryProvider(MemoryProvider):
         # "local" is a legacy alias for "local_embedded"
         if self._mode == "local":
             self._mode = "local_embedded"
+        if self._mode == "local_embedded" and not self._config.get("profile"):
+            self._config["profile"] = _derive_local_profile_name(
+                agent_identity=agent_identity,
+                hermes_home=hermes_home,
+            )
         self._api_key = self._config.get("apiKey") or self._config.get("api_key") or os.environ.get("HINDSIGHT_API_KEY", "")
         default_url = _DEFAULT_LOCAL_URL if self._mode in ("local_embedded", "local_external") else _DEFAULT_API_URL
         self._api_url = self._config.get("api_url") or os.environ.get("HINDSIGHT_API_URL", default_url)

run_agent.py

@@ -604,6 +604,8 @@ class AIAgent:
         checkpoint_max_snapshots: int = 50,
         pass_session_id: bool = False,
         persist_session: bool = True,
+        context_snapshots_enabled: bool | None = None,
+        context_snapshots_dir: str | None = None,
     ):
         """
         Initialize the AI Agent.
@@ -1129,6 +1131,43 @@ class AIAgent:
         except Exception:
             _agent_cfg = {}
 
+        def _is_enabled(value):
+            if isinstance(value, bool):
+                return value
+            return str(value).strip().lower() in {"1", "true", "yes", "on"}
+
+        _debug_cfg = _agent_cfg.get("debug", {}) if isinstance(_agent_cfg, dict) else {}
+        if not isinstance(_debug_cfg, dict):
+            _debug_cfg = {}
+        _snapshot_cfg = _debug_cfg.get("context_snapshots", {})
+        if not isinstance(_snapshot_cfg, dict):
+            _snapshot_cfg = {}
+        _snapshots_env = os.getenv("HERMES_CONTEXT_SNAPSHOTS")
+        _snapshots_dir_env = os.getenv("HERMES_CONTEXT_SNAPSHOTS_DIR")
+        if context_snapshots_enabled is None:
+            if _snapshots_env is not None:
+                self._context_snapshots_enabled = _is_enabled(_snapshots_env)
+            else:
+                self._context_snapshots_enabled = _is_enabled(_snapshot_cfg.get("enabled", False))
+        else:
+            self._context_snapshots_enabled = bool(context_snapshots_enabled)
+        self._context_snapshots_dir = (
+            context_snapshots_dir
+            or _snapshots_dir_env
+            or _snapshot_cfg.get("dir")
+            or None
+        )
+        try:
+            from agent.context_snapshots import ContextSnapshotRecorder
+            self._context_snapshot_recorder = ContextSnapshotRecorder(
+                session_id=self.session_id,
+                enabled=self._context_snapshots_enabled,
+                base_dir=self._context_snapshots_dir,
+            )
+        except Exception as _snapshot_err:
+            logger.debug("Context snapshot recorder init failed: %s", _snapshot_err)
+            self._context_snapshot_recorder = None
+
         # Persistent memory (MEMORY.md + USER.md) -- loaded from disk
         self._memory_store = None
         self._memory_enabled = False
@@ -8144,12 +8183,17 @@ class AIAgent:
         # Use original_user_message (clean input) — user_message may contain
         # injected skill content that bloats / breaks provider queries.
         _ext_prefetch_cache = ""
+        _memory_provider_prompt_cache = ""
         if self._memory_manager:
             try:
                 _query = original_user_message if isinstance(original_user_message, str) else ""
                 _ext_prefetch_cache = self._memory_manager.prefetch_all(_query) or ""
             except Exception:
                 pass
+            try:
+                _memory_provider_prompt_cache = self._memory_manager.build_system_prompt() or ""
+            except Exception:
+                pass
 
         while (api_call_count < self.max_iterations and self.iteration_budget.remaining > 0) or self._budget_grace_call:
             # Reset per-turn checkpoint dedup so each iteration can take one snapshot
@@ -8217,6 +8261,8 @@ class AIAgent:
             # However, providers like Moonshot AI require a separate 'reasoning_content' field
             # on assistant messages with tool_calls. We handle both cases here.
             api_messages = []
+            _current_api_user_message = ""
+            _current_memory_context_block = ""
             for idx, msg in enumerate(messages):
                 api_msg = msg.copy()
 
@@ -8231,12 +8277,15 @@ class AIAgent:
                         _fenced = build_memory_context_block(_ext_prefetch_cache)
                         if _fenced:
                             _injections.append(_fenced)
+                            _current_memory_context_block = _fenced
                     if _plugin_user_context:
                         _injections.append(_plugin_user_context)
                     if _injections:
                         _base = api_msg.get("content", "")
                         if isinstance(_base, str):
                             api_msg["content"] = _base + "\n\n" + "\n\n".join(_injections)
+                    if isinstance(api_msg.get("content"), str):
+                        _current_api_user_message = api_msg["content"]
 
                 # For ALL assistant messages, pass reasoning back to the API
                 # This ensures multi-turn reasoning context is preserved
@@ -8271,7 +8320,13 @@ class AIAgent:
                 from agent.privacy_filter import PrivacyFilter
                 pf = PrivacyFilter()
                 # Sanitize messages before they reach the provider
-                api_messages = pf.sanitize_messages(api_messages)
+                _pf_result = pf.sanitize_messages(api_messages)
+                if isinstance(_pf_result, tuple):
+                    api_messages, _pf_report = _pf_result
+                    if getattr(pf, "last_report", None) is None:
+                        pf.last_report = _pf_report
+                else:
+                    api_messages = _pf_result
                 if pf.last_report and pf.last_report.had_redactions:
                     logger.info(f"Privacy Filter: Redacted sensitive data from turn payload. Details: {pf.last_report.summary()}")
             except Exception as e:
@@ -8342,6 +8397,27 @@ class AIAgent:
                     new_tcs.append(tc)
                 am["tool_calls"] = new_tcs
 
+            if self._context_snapshot_recorder:
+                try:
+                    self._context_snapshot_recorder.record_call(
+                        api_call_count,
+                        system_prompt=effective_system,
+                        memory_provider_system_prompt=_memory_provider_prompt_cache,
+                        memory_prefetch_raw=_ext_prefetch_cache,
+                        memory_context_block=_current_memory_context_block,
+                        api_user_message=_current_api_user_message,
+                        api_messages=api_messages,
+                        metadata={
+                            "model": self.model,
+                            "provider": self.provider,
+                            "platform": self.platform or "",
+                            "api_mode": self.api_mode,
+                            "memory_providers": [p.name for p in getattr(self._memory_manager, "providers", [])],
+                        },
+                    )
+                except Exception as _snapshot_err:
+                    logger.debug("Context snapshot capture failed: %s", _snapshot_err)
+
             # Calculate approximate request size for logging
             total_chars = sum(len(str(msg)) for msg in api_messages)
             approx_tokens = estimate_messages_tokens_rough(api_messages)

tests/agent/test_context_snapshots.py

@@ -0,0 +1,43 @@
+from pathlib import Path
+
+from agent.context_snapshots import ContextSnapshotRecorder
+
+
+def test_disabled_recorder_writes_nothing(tmp_path):
+    recorder = ContextSnapshotRecorder(session_id="session-1", enabled=False, base_dir=tmp_path)
+
+    out = recorder.record_call(
+        1,
+        system_prompt="system",
+        api_messages=[{"role": "user", "content": "hello"}],
+    )
+
+    assert out is None
+    assert not (tmp_path / "session-1").exists()
+
+
+def test_enabled_recorder_writes_expected_artifacts(tmp_path):
+    recorder = ContextSnapshotRecorder(session_id="session-1", enabled=True, base_dir=tmp_path)
+
+    out = recorder.record_call(
+        1,
+        system_prompt="system prompt",
+        memory_provider_system_prompt="# Hindsight Memory\nActive.",
+        memory_prefetch_raw="- remembered fact",
+        memory_context_block="<memory-context>\nremembered\n</memory-context>",
+        api_user_message="What do I prefer?\n\n<memory-context>\nremembered\n</memory-context>",
+        api_messages=[
+            {"role": "system", "content": "system prompt"},
+            {"role": "user", "content": "What do I prefer?"},
+        ],
+        metadata={"provider": "openai", "memory_providers": ["builtin", "hindsight"]},
+    )
+
+    assert out == tmp_path / "session-1" / "call_001"
+    assert (out / "system_prompt.txt").read_text(encoding="utf-8") == "system prompt"
+    assert (out / "memory_provider_system_prompt.txt").read_text(encoding="utf-8").startswith("# Hindsight Memory")
+    assert (out / "memory_prefetch_raw.txt").read_text(encoding="utf-8") == "- remembered fact"
+    assert "<memory-context>" in (out / "memory_context_block.txt").read_text(encoding="utf-8")
+    assert "What do I prefer?" in (out / "api_user_message.txt").read_text(encoding="utf-8")
+    assert (out / "api_messages.json").read_text(encoding="utf-8").startswith("[")
+    assert '"hindsight"' in (out / "metadata.json").read_text(encoding="utf-8")

tests/plugins/memory/test_hindsight_provider.py

@@ -596,3 +596,26 @@ class TestAvailability:
         monkeypatch.setenv("HINDSIGHT_MODE", "local")
         p = HindsightMemoryProvider()
         assert p.is_available()
+
+    def test_local_embedded_profile_defaults_to_agent_identity(self, tmp_path, monkeypatch):
+        config_path = tmp_path / "hindsight" / "config.json"
+        config_path.parent.mkdir(parents=True, exist_ok=True)
+        config_path.write_text(json.dumps({
+            "mode": "local_embedded",
+            "llm_provider": "ollama",
+            "llm_model": "gemma3:12b",
+        }))
+        monkeypatch.setattr(
+            "plugins.memory.hindsight.get_hermes_home",
+            lambda: tmp_path,
+        )
+
+        p = HindsightMemoryProvider()
+        p.initialize(
+            session_id="test-session",
+            hermes_home=str(tmp_path / "profiles" / "atlas-hindsight"),
+            platform="cli",
+            agent_identity="atlas-hindsight",
+        )
+
+        assert p._config["profile"] == "atlas-hindsight"

tests/run_agent/test_run_agent_context_snapshots.py

@@ -0,0 +1,94 @@
+from pathlib import Path
+from types import SimpleNamespace
+from unittest.mock import MagicMock, patch
+import importlib
+import sys
+import types
+
+
+
+def _make_tool_defs(*names: str) -> list:
+    return [
+        {
+            "type": "function",
+            "function": {
+                "name": n,
+                "description": f"{n} tool",
+                "parameters": {"type": "object", "properties": {}},
+            },
+        }
+        for n in names
+    ]
+
+
+
+def _mock_response(content="Done", finish_reason="stop"):
+    msg = SimpleNamespace(content=content, tool_calls=None)
+    choice = SimpleNamespace(message=msg, finish_reason=finish_reason)
+    return SimpleNamespace(choices=[choice], usage=SimpleNamespace(prompt_tokens=1, completion_tokens=1, total_tokens=2))
+
+
+
+def _load_ai_agent():
+    sys.modules.setdefault("agent.auxiliary_client", types.SimpleNamespace(call_llm=lambda *a, **k: ""))
+    run_agent = importlib.import_module("run_agent")
+    return run_agent.AIAgent
+
+
+
+def test_run_conversation_writes_context_snapshot_artifacts(tmp_path):
+    AIAgent = _load_ai_agent()
+
+    class _FakePrivacyFilter:
+        def __init__(self):
+            self.last_report = None
+
+        def sanitize_messages(self, messages):
+            return list(messages)
+
+    with (
+        patch("run_agent.get_tool_definitions", return_value=_make_tool_defs("web_search")),
+        patch("run_agent.check_toolset_requirements", return_value={}),
+        patch("run_agent.OpenAI"),
+        patch("hermes_cli.plugins.invoke_hook", return_value=[]),
+        patch.dict(sys.modules, {"agent.privacy_filter": types.SimpleNamespace(PrivacyFilter=_FakePrivacyFilter)}),
+    ):
+        agent = AIAgent(
+            api_key="test-key-1234567890",
+            base_url="https://example.com/v1",
+            quiet_mode=True,
+            skip_context_files=True,
+            skip_memory=True,
+            context_snapshots_enabled=True,
+            context_snapshots_dir=str(tmp_path),
+        )
+
+    agent.client = MagicMock()
+    agent.client.chat.completions.create.return_value = _mock_response(content="Done")
+    agent._build_system_prompt = MagicMock(return_value="Core system prompt")
+    agent._memory_manager = MagicMock()
+    agent._memory_manager.prefetch_all.return_value = "- remembered preference"
+    agent._memory_manager.build_system_prompt.return_value = "# Hindsight Memory\nActive."
+    agent._memory_manager.providers = [
+        SimpleNamespace(name="builtin"),
+        SimpleNamespace(name="hindsight"),
+    ]
+
+    result = agent.run_conversation("What do I prefer?")
+
+    assert result["final_response"] == "Done"
+
+    call_dir = tmp_path / agent.session_id / "call_001"
+    assert call_dir.exists()
+    assert (call_dir / "system_prompt.txt").read_text(encoding="utf-8") == "Core system prompt"
+    assert (call_dir / "memory_provider_system_prompt.txt").read_text(encoding="utf-8").startswith("# Hindsight Memory")
+    assert (call_dir / "memory_prefetch_raw.txt").read_text(encoding="utf-8") == "- remembered preference"
+    assert "<memory-context>" in (call_dir / "memory_context_block.txt").read_text(encoding="utf-8")
+    api_user_message = (call_dir / "api_user_message.txt").read_text(encoding="utf-8")
+    assert "What do I prefer?" in api_user_message
+    assert "remembered preference" in api_user_message
+    api_messages = (call_dir / "api_messages.json").read_text(encoding="utf-8")
+    assert '"role": "system"' in api_messages
+    assert '"role": "user"' in api_messages
+    metadata = (call_dir / "metadata.json").read_text(encoding="utf-8")
+    assert '"hindsight"' in metadata

tests/test_hindsight_local_eval_doc.py

@@ -0,0 +1,22 @@
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+DOC = ROOT / "docs" / "hindsight-local-eval.md"
+
+
+def test_hindsight_local_eval_doc_exists_and_covers_switching():
+    assert DOC.exists(), "missing Hindsight local eval doc"
+    text = DOC.read_text(encoding="utf-8")
+    for snippet in (
+        "atlas-baseline",
+        "atlas-mempalace",
+        "atlas-hindsight",
+        "HERMES_HOME",
+        "HERMES_CONTEXT_SNAPSHOTS",
+        "memory_prefetch_raw.txt",
+        "api_user_message.txt",
+        "local_embedded",
+        "hindsight/config.json",
+    ):
+        assert snippet in text

tests/tools/test_tts_kittentts.py

@@ -1,236 +0,0 @@
-"""Tests for the KittenTTS local provider in tools/tts_tool.py."""
-
-import json
-from unittest.mock import MagicMock, patch
-
-import numpy as np
-import pytest
-
-
-@pytest.fixture(autouse=True)
-def clean_env(monkeypatch):
-    for key in ("HERMES_SESSION_PLATFORM",):
-        monkeypatch.delenv(key, raising=False)
-
-
-@pytest.fixture(autouse=True)
-def clear_kittentts_cache():
-    """Reset the module-level model cache between tests."""
-    from tools import tts_tool as _tt
-    _tt._kittentts_model_cache.clear()
-    yield
-    _tt._kittentts_model_cache.clear()
-
-
-@pytest.fixture
-def mock_kittentts_module():
-    """Inject a fake kittentts + soundfile module that return stub objects."""
-    fake_model = MagicMock()
-    # 24kHz float32 PCM at ~2s of silence
-    fake_model.generate.return_value = np.zeros(48000, dtype=np.float32)
-    fake_cls = MagicMock(return_value=fake_model)
-    fake_kittentts = MagicMock()
-    fake_kittentts.KittenTTS = fake_cls
-
-    # Stub soundfile — the real package isn't installed in CI venv, and
-    # _generate_kittentts does `import soundfile as sf` at runtime.
-    fake_sf = MagicMock()
-
-    def _fake_write(path, audio, samplerate):
-        # Emulate writing a real file so downstream path checks succeed.
-        import pathlib
-
-        pathlib.Path(path).write_bytes(b"RIFF\x00\x00\x00\x00WAVEfmt fake")
-
-    fake_sf.write = _fake_write
-
-    with patch.dict(
-        "sys.modules",
-        {"kittentts": fake_kittentts, "soundfile": fake_sf},
-    ):
-        yield fake_model, fake_cls
-
-
-class TestGenerateKittenTts:
-    def test_successful_wav_generation(self, tmp_path, mock_kittentts_module):
-        from tools.tts_tool import _generate_kittentts
-
-        fake_model, fake_cls = mock_kittentts_module
-        output_path = str(tmp_path / "test.wav")
-        result = _generate_kittentts("Hello world", output_path, {})
-
-        assert result == output_path
-        assert (tmp_path / "test.wav").exists()
-        fake_cls.assert_called_once()
-        fake_model.generate.assert_called_once()
-
-    def test_config_passes_voice_speed_cleantext(self, tmp_path, mock_kittentts_module):
-        from tools.tts_tool import _generate_kittentts
-
-        fake_model, _ = mock_kittentts_module
-        config = {
-            "kittentts": {
-                "model": "KittenML/kitten-tts-mini-0.8",
-                "voice": "Luna",
-                "speed": 1.25,
-                "clean_text": False,
-            }
-        }
-        _generate_kittentts("Hi there", str(tmp_path / "out.wav"), config)
-
-        call_kwargs = fake_model.generate.call_args.kwargs
-        assert call_kwargs["voice"] == "Luna"
-        assert call_kwargs["speed"] == 1.25
-        assert call_kwargs["clean_text"] is False
-
-    def test_default_model_and_voice(self, tmp_path, mock_kittentts_module):
-        from tools.tts_tool import (
-            DEFAULT_KITTENTTS_MODEL,
-            DEFAULT_KITTENTTS_VOICE,
-            _generate_kittentts,
-        )
-
-        fake_model, fake_cls = mock_kittentts_module
-        _generate_kittentts("Hi", str(tmp_path / "out.wav"), {})
-
-        fake_cls.assert_called_once_with(DEFAULT_KITTENTTS_MODEL)
-        assert fake_model.generate.call_args.kwargs["voice"] == DEFAULT_KITTENTTS_VOICE
-
-    def test_model_is_cached_across_calls(self, tmp_path, mock_kittentts_module):
-        from tools.tts_tool import _generate_kittentts
-
-        _, fake_cls = mock_kittentts_module
-        _generate_kittentts("One", str(tmp_path / "a.wav"), {})
-        _generate_kittentts("Two", str(tmp_path / "b.wav"), {})
-
-        # Same model name → class instantiated exactly once
-        assert fake_cls.call_count == 1
-
-    def test_different_models_are_cached_separately(self, tmp_path, mock_kittentts_module):
-        from tools.tts_tool import _generate_kittentts
-
-        _, fake_cls = mock_kittentts_module
-        _generate_kittentts(
-            "A",
-            str(tmp_path / "a.wav"),
-            {"kittentts": {"model": "KittenML/kitten-tts-nano-0.8-int8"}},
-        )
-        _generate_kittentts(
-            "B",
-            str(tmp_path / "b.wav"),
-            {"kittentts": {"model": "KittenML/kitten-tts-mini-0.8"}},
-        )
-
-        assert fake_cls.call_count == 2
-
-    def test_non_wav_extension_triggers_ffmpeg_conversion(
-        self, tmp_path, mock_kittentts_module, monkeypatch
-    ):
-        """Non-.wav output path causes WAV → target ffmpeg conversion."""
-        from tools import tts_tool as _tt
-
-        calls = []
-
-        def fake_shutil_which(cmd):
-            return "/usr/bin/ffmpeg" if cmd == "ffmpeg" else None
-
-        def fake_run(cmd, check=False, timeout=None, **kw):
-            calls.append(cmd)
-            # Emulate ffmpeg writing the output file
-            import pathlib
-
-            out_path = cmd[-1]
-            pathlib.Path(out_path).write_bytes(b"fake-mp3-data")
-            return MagicMock(returncode=0)
-
-        monkeypatch.setattr(_tt.shutil, "which", fake_shutil_which)
-        monkeypatch.setattr(_tt.subprocess, "run", fake_run)
-
-        output_path = str(tmp_path / "test.mp3")
-        result = _tt._generate_kittentts("Hi", output_path, {})
-
-        assert result == output_path
-        assert len(calls) == 1
-        assert calls[0][0] == "/usr/bin/ffmpeg"
-
-    def test_missing_kittentts_raises_import_error(self, tmp_path, monkeypatch):
-        """When kittentts package is not installed, _import_kittentts raises."""
-        import sys
-
-        monkeypatch.setitem(sys.modules, "kittentts", None)
-        from tools.tts_tool import _generate_kittentts
-
-        with pytest.raises((ImportError, TypeError)):
-            _generate_kittentts("Hi", str(tmp_path / "out.wav"), {})
-
-
-class TestCheckKittenttsAvailable:
-    def test_reports_available_when_package_present(self, monkeypatch):
-        import importlib.util
-        from tools.tts_tool import _check_kittentts_available
-
-        fake_spec = MagicMock()
-        monkeypatch.setattr(
-            importlib.util,
-            "find_spec",
-            lambda name: fake_spec if name == "kittentts" else None,
-        )
-        assert _check_kittentts_available() is True
-
-    def test_reports_unavailable_when_package_missing(self, monkeypatch):
-        import importlib.util
-        from tools.tts_tool import _check_kittentts_available
-
-        monkeypatch.setattr(importlib.util, "find_spec", lambda name: None)
-        assert _check_kittentts_available() is False
-
-
-class TestDispatcherBranch:
-    def test_kittentts_not_installed_returns_helpful_error(self, monkeypatch, tmp_path):
-        """When provider=kittentts but package missing, return JSON error with setup hint."""
-        import sys
-
-        monkeypatch.setitem(sys.modules, "kittentts", None)
-        monkeypatch.setenv("HERMES_HOME", str(tmp_path))
-
-        from tools.tts_tool import text_to_speech_tool
-
-        # Write a config telling it to use kittentts
-        import yaml
-
-        (tmp_path / "config.yaml").write_text(
-            yaml.safe_dump({"tts": {"provider": "kittentts"}})
-        )
-
-        result = json.loads(text_to_speech_tool(text="Hello"))
-        assert result["success"] is False
-        assert "kittentts" in result["error"].lower()
-        assert "hermes setup tts" in result["error"].lower()
-
-    def test_non_telegram_explicit_wav_path_is_preserved(
-        self, monkeypatch, tmp_path, mock_kittentts_module
-    ):
-        """Explicit WAV outputs should stay WAV outside Telegram sessions."""
-        import yaml
-        from tools import tts_tool as _tt
-
-        monkeypatch.setenv("HERMES_HOME", str(tmp_path))
-        (tmp_path / "config.yaml").write_text(
-            yaml.safe_dump({"tts": {"provider": "kittentts"}})
-        )
-
-        def fail_convert(_path):
-            raise AssertionError("_convert_to_opus should not run outside Telegram")
-
-        monkeypatch.setattr(_tt, "_convert_to_opus", fail_convert)
-
-        result = json.loads(
-            _tt.text_to_speech_tool(
-                text="Hello from KittenTTS",
-                output_path=str(tmp_path / "out.wav"),
-            )
-        )
-
-        assert result["success"] is True
-        assert result["file_path"] == str(tmp_path / "out.wav")
-        assert (tmp_path / "out.wav").exists()

tools/tts_tool.py

@@ -2,14 +2,13 @@
 """
 Text-to-Speech Tool Module
 
-Supports seven TTS providers:
+Supports six TTS providers:
 - Edge TTS (default, free, no API key): Microsoft Edge neural voices
 - ElevenLabs (premium): High-quality voices, needs ELEVENLABS_API_KEY
 - OpenAI TTS: Good quality, needs OPENAI_API_KEY
 - MiniMax TTS: High-quality with voice cloning, needs MINIMAX_API_KEY
 - Mistral (Voxtral TTS): Multilingual, native Opus, needs MISTRAL_API_KEY
 - NeuTTS (local, free, no API key): On-device TTS via neutts_cli, needs neutts installed
-- KittenTTS (local, free, no API key): Lightweight on-device ONNX TTS via kittentts
 
 Output formats:
 - Opus (.ogg) for Telegram voice bubbles (requires ffmpeg for Edge TTS)
@@ -78,12 +77,6 @@ def _import_sounddevice():
     return sd
 
 
-def _import_kittentts():
-    """Lazy import KittenTTS. Returns the class or raises ImportError."""
-    from kittentts import KittenTTS
-    return KittenTTS
-
-
 # ===========================================================================
 # Defaults
 # ===========================================================================
@@ -93,8 +86,6 @@ DEFAULT_ELEVENLABS_VOICE_ID = "pNInz6obpgDQGcFmaJgB"  # Adam
 DEFAULT_ELEVENLABS_MODEL_ID = "eleven_multilingual_v2"
 DEFAULT_ELEVENLABS_STREAMING_MODEL_ID = "eleven_flash_v2_5"
 DEFAULT_OPENAI_MODEL = "gpt-4o-mini-tts"
-DEFAULT_KITTENTTS_MODEL = "KittenML/kitten-tts-nano-0.8-int8"  # 25MB
-DEFAULT_KITTENTTS_VOICE = "Jasper"
 DEFAULT_OPENAI_VOICE = "alloy"
 DEFAULT_OPENAI_BASE_URL = "https://api.openai.com/v1"
 DEFAULT_MINIMAX_MODEL = "speech-2.8-hd"
@@ -457,15 +448,6 @@ def _check_neutts_available() -> bool:
         return False
 
 
-def _check_kittentts_available() -> bool:
-    """Check if the kittentts engine is importable (installed locally)."""
-    try:
-        import importlib.util
-        return importlib.util.find_spec("kittentts") is not None
-    except Exception:
-        return False
-
-
 def _default_neutts_ref_audio() -> str:
     """Return path to the bundled default voice reference audio."""
     return str(Path(__file__).parent / "neutts_samples" / "jo.wav")
@@ -529,51 +511,6 @@ def _generate_neutts(text: str, output_path: str, tts_config: Dict[str, Any]) ->
     return output_path
 
 
-# ===========================================================================
-# Provider: KittenTTS (local, lightweight)
-# ===========================================================================
-
-# Module-level cache for KittenTTS model instances
-_kittentts_model_cache: Dict[str, Any] = {}
-
-
-def _generate_kittentts(text: str, output_path: str, tts_config: Dict[str, Any]) -> str:
-    """Generate speech using the local KittenTTS ONNX model."""
-    KittenTTS = _import_kittentts()
-    kt_config = tts_config.get("kittentts", {})
-    model_name = kt_config.get("model", DEFAULT_KITTENTTS_MODEL)
-    voice = kt_config.get("voice", DEFAULT_KITTENTTS_VOICE)
-    speed = kt_config.get("speed", 1.0)
-    clean_text = kt_config.get("clean_text", True)
-
-    global _kittentts_model_cache
-    if model_name not in _kittentts_model_cache:
-        logger.info("[KittenTTS] Loading model: %s", model_name)
-        _kittentts_model_cache[model_name] = KittenTTS(model_name)
-
-    model = _kittentts_model_cache[model_name]
-    audio = model.generate(text, voice=voice, speed=speed, clean_text=clean_text)
-
-    import soundfile as sf
-
-    wav_path = output_path
-    if not output_path.endswith(".wav"):
-        wav_path = output_path.rsplit(".", 1)[0] + ".wav"
-
-    sf.write(wav_path, audio, 24000)
-
-    if wav_path != output_path:
-        ffmpeg = shutil.which("ffmpeg")
-        if ffmpeg:
-            conv_cmd = [ffmpeg, "-i", wav_path, "-y", "-loglevel", "error", output_path]
-            subprocess.run(conv_cmd, check=True, timeout=30)
-            os.remove(wav_path)
-        else:
-            os.rename(wav_path, output_path)
-
-    return output_path
-
-
 # ===========================================================================
 # Main tool function
 # ===========================================================================
@@ -685,19 +622,6 @@ def text_to_speech_tool(
             logger.info("Generating speech with NeuTTS (local)...")
             _generate_neutts(text, file_str, tts_config)
 
-        elif provider == "kittentts":
-            try:
-                _import_kittentts()
-            except ImportError:
-                return json.dumps({
-                    "success": False,
-                    "error": "KittenTTS provider selected but 'kittentts' package not installed. "
-                             "Run 'hermes setup tts' and choose KittenTTS, or install manually: "
-                             "pip install https://github.com/KittenML/KittenTTS/releases/download/0.8.1/kittentts-0.8.1-py3-none-any.whl"
-                }, ensure_ascii=False)
-            logger.info("Generating speech with KittenTTS (local, lightweight)...")
-            _generate_kittentts(text, file_str, tts_config)
-
         else:
             # Default: Edge TTS (free), with NeuTTS as local fallback
             edge_available = True
@@ -734,10 +658,10 @@ def text_to_speech_tool(
                 "error": f"TTS generation produced no output (provider: {provider})"
             }, ensure_ascii=False)
 
-        # Try Opus conversion for Telegram compatibility only.
-        # Outside Telegram, preserve the caller's explicit output format.
+        # Try Opus conversion for Telegram compatibility
+        # Edge TTS outputs MP3, NeuTTS outputs WAV — both need ffmpeg conversion
         voice_compatible = False
-        if want_opus and provider in ("edge", "neutts", "minimax", "kittentts") and not file_str.endswith(".ogg"):
+        if provider in ("edge", "neutts", "minimax") and not file_str.endswith(".ogg"):
             opus_path = _convert_to_opus(file_str)
             if opus_path:
                 file_str = opus_path
@@ -818,8 +742,6 @@ def check_tts_requirements() -> bool:
         pass
     if _check_neutts_available():
         return True
-    if _check_kittentts_available():
-        return True
     return False
 
 

website/docs/user-guide/features/tts.md

@@ -10,7 +10,7 @@ Hermes Agent supports both text-to-speech output and voice message transcription
 
 ## Text-to-Speech
 
-Convert text to speech with seven providers:
+Convert text to speech with six providers:
 
 | Provider | Quality | Cost | API Key |
 |----------|---------|------|---------|
@@ -20,7 +20,6 @@ Convert text to speech with seven providers:
 | **MiniMax TTS** | Excellent | Paid | `MINIMAX_API_KEY` |
 | **Mistral (Voxtral TTS)** | Excellent | Paid | `MISTRAL_API_KEY` |
 | **NeuTTS** | Good | Free | None needed |
-| **KittenTTS** | Good | Free (local) | None needed |
 
 ### Platform Delivery
 
@@ -36,7 +35,7 @@ Convert text to speech with seven providers:
 ```yaml
 # In ~/.hermes/config.yaml
 tts:
-  provider: "edge"              # "edge" | "elevenlabs" | "openai" | "minimax" | "mistral" | "neutts" | "kittentts"
+  provider: "edge"              # "edge" | "elevenlabs" | "openai" | "minimax" | "mistral" | "neutts"
   speed: 1.0                    # Global speed multiplier (provider-specific settings override this)
   edge:
     voice: "en-US-AriaNeural"   # 322 voices, 74 languages
@@ -63,11 +62,6 @@ tts:
     ref_text: ''
     model: neuphonic/neutts-air-q4-gguf
     device: cpu
-  kittentts:
-    model: KittenML/kitten-tts-nano-0.8-int8   # 25MB int8 default; also micro and mini variants
-    voice: Jasper                               # Jasper, Bella, Luna, Bruno, Rosie, Hugo, Kiki, Leo
-    speed: 1.0
-    clean_text: true
 ```
 
 **Speed control**: The global `tts.speed` value applies to all providers by default. Each provider can override it with its own `speed` setting (e.g., `tts.openai.speed: 1.5`). Provider-specific speed takes precedence over the global value. Default is `1.0` (normal speed).
@@ -80,7 +74,6 @@ Telegram voice bubbles require Opus/OGG audio format:
 - **Edge TTS** (default) outputs MP3 and needs **ffmpeg** to convert:
 - **MiniMax TTS** outputs MP3 and needs **ffmpeg** to convert for Telegram voice bubbles
 - **NeuTTS** outputs WAV and also needs **ffmpeg** to convert for Telegram voice bubbles
-- **KittenTTS** outputs WAV and also needs **ffmpeg** to convert for Telegram voice bubbles
 
 ```bash
 # Ubuntu/Debian
@@ -93,7 +86,7 @@ brew install ffmpeg
 sudo dnf install ffmpeg
 ```
 
-Without ffmpeg, Edge TTS, MiniMax TTS, NeuTTS, and KittenTTS audio are sent as regular audio files (playable, but shown as a rectangular player instead of a voice bubble).
+Without ffmpeg, Edge TTS, MiniMax TTS, and NeuTTS audio are sent as regular audio files (playable, but shown as a rectangular player instead of a voice bubble).
 
 :::tip
 If you want voice bubbles without installing ffmpeg, switch to the OpenAI, ElevenLabs, or Mistral provider.