Sankhya-AI · Ashish-dwi99 · Mar 31, 2026 · Mar 31, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -4,6 +4,46 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/), and this project adheres to [Semantic Versioning](https://semver.org/).
 
+## [2.2.0b1] - 2026-03-31 — Architectural Cleanup
+
+Beta release focused on internal discipline rather than new features.
+
+### Changed — Architecture
+
+- **main.py decomposition**: Extracted `SearchPipeline`, `MemoryWritePipeline`, `OrchestrationEngine` from the 6,129-line monolith. main.py is now ~3,100 lines (49% reduction).
+- **Public surface**: `dhee/__init__.py` rewritten for clean, narrow exports. `Memory = CoreMemory` (not FullMemory). Cognitive subsystems intentionally kept internal.
+- **MCP split**: `mcp_slim.py` (4-tool product surface) vs `mcp_server.py` (24-tool power surface). Clear separation of concerns.
+
+### Changed — Rename Debt
+
+- All `FADEM_*` env vars → `DHEE_*` (with `FADEM_*` fallback for backward compat).
+- Internal `fadem_config` → `fade_config` across memory package.
+- Default collection name `fadem_memories` → `dhee_memories`.
+- Config field `MemoryConfig.engram` → `MemoryConfig.fade`.
+- CLI, MCP server, observability, presets: all `engram` product references removed.
+
+### Added — D2Skill Policy Improvements
+
+- **Dual-granularity policies**: `PolicyGranularity.TASK` (strategy) vs `PolicyGranularity.STEP` (local correction). Inspired by D2Skill (arXiv:2603.28716).
+- **Utility scoring**: EMA-smoothed performance delta tracking on policies. Three-signal retrieval ranking (condition match + sigmoid utility + UCB exploration bonus).
+- **Utility-based pruning**: `PolicyStore.prune()` removes deprecated policies first, protects validated ones.
+- **Buddhi wiring**: `reflect()` accepts `outcome_score`, computes baseline vs actual delta, feeds it to policy `record_outcome()`.
+
+### Fixed
+
+- `buddhi.py`: Replaced dead `memory.get_last_session_digest()` call with working `get_last_session()` import.
+- `mcp_server.py`: Fixed wrong "8 tools total" comment (actually 24), fixed `-> Memory` type hints (Memory not imported).
+- CLI: Wired `benchmark` command into parser (existed but was unreachable).
+- `PolicyStore.prune()`: Fixed bug where `candidates.pop()` pruned validated policies instead of deprecated ones.
+- CHANGELOG 2.1.0: Removed false "Production/Stable" and "A-grade" claims.
+
+### Changed — Packaging
+
+- Version: 2.1.0 → 2.2.0b1
+- Classifier: `Development Status :: 4 - Beta` (was falsely claiming Production/Stable)
+
+---
+
 ## [2.1.0] - 2026-03-30 — Cognition Primitives
 
 Dhee V2.1: Adds first-class cognitive primitives (episodes, tasks, policies, beliefs, triggers) and a 60-test suite that exercises them. These are internal building blocks — the public API remains the 4-operation surface (remember/recall/context/checkpoint).

diff --git a/dhee/cli.py b/dhee/cli.py
@@ -10,6 +10,7 @@
     dhee categories         List categories
     dhee export             Export to JSON
     dhee import <file>      Import from JSON
+    dhee benchmark          Run performance benchmarks
     dhee status             Version, config, DB info
 """
 
@@ -457,6 +458,9 @@ def build_parser() -> argparse.ArgumentParser:
     p_status = sub.add_parser("status", help="Show version, config, and agents")
     p_status.add_argument("--json", action="store_true", help="JSON output")
 
+    # benchmark
+    sub.add_parser("benchmark", help="Run performance benchmarks")
+
     # uninstall
     sub.add_parser("uninstall", help="Remove ~/.dhee directory")
 
@@ -477,6 +481,7 @@ def build_parser() -> argparse.ArgumentParser:
     "export": cmd_export,
     "import": cmd_import,
     "status": cmd_status,
+    "benchmark": cmd_benchmark,
     "uninstall": cmd_uninstall,
 }
 

diff --git a/dhee/cli_config.py b/dhee/cli_config.py
@@ -125,15 +125,15 @@ def get_memory_instance(config: Optional[Dict[str, Any]] = None):
             provider="sqlite_vec",
             config={
                 "path": vec_db_path,
-                "collection_name": "fadem_memories",
+                "collection_name": "dhee_memories",
                 "embedding_model_dims": embedding_dims,
             },
         ),
         llm=LLMConfig(provider=provider, config=llm_cfg),
         embedder=EmbedderConfig(provider=provider, config=embedder_cfg),
         history_db_path=history_db_path,
         embedding_model_dims=embedding_dims,
-        engram=FadeMemConfig(enable_forgetting=True),
+        fade=FadeMemConfig(enable_forgetting=True),
     )
 
     return FullMemory(memory_config)
diff --git a/dhee/configs/base.py b/dhee/configs/base.py
@@ -24,7 +24,7 @@ class VectorStoreConfig(BaseModel):
     config: Dict[str, Any] = Field(
         default_factory=lambda: {
             "path": os.path.join(_dhee_data_dir(), "zvec"),
-            "collection_name": "fadem_memories",
+            "collection_name": "dhee_memories",
         }
     )
 
@@ -605,15 +605,15 @@ class MemoryConfig(BaseModel):
     history_db_path: str = Field(
         default_factory=lambda: os.path.join(_dhee_data_dir(), "history.db")
     )
-    collection_name: str = "fadem_memories"
+    collection_name: str = "dhee_memories"
     embedding_model_dims: int = 4096  # nvidia/nv-embed-v1 default dimensions
     version: str = "v1.4"  # Updated for CLS Distillation Memory
     custom_fact_extraction_prompt: Optional[str] = None
     custom_conflict_prompt: Optional[str] = None
     custom_fusion_prompt: Optional[str] = None
     custom_echo_prompt: Optional[str] = None
     custom_category_prompt: Optional[str] = None
-    engram: FadeMemConfig = Field(default_factory=FadeMemConfig)
+    fade: FadeMemConfig = Field(default_factory=FadeMemConfig)
     echo: EchoMemConfig = Field(default_factory=EchoMemConfig)
     category: CategoryMemConfig = Field(default_factory=CategoryMemConfig)
     scope: ScopeConfig = Field(default_factory=ScopeConfig)

diff --git a/dhee/configs/presets.py b/dhee/configs/presets.py
@@ -45,7 +45,7 @@ def minimal_config():
         history_db_path=os.path.join(data_dir, "history.db"),
         collection_name="dhee_memories",
         embedding_model_dims=384,
-        engram=FadeMemConfig(enable_forgetting=True),
+        fade=FadeMemConfig(enable_forgetting=True),
         echo=EchoMemConfig(enable_echo=False),
         category=CategoryMemConfig(enable_categories=False),
         graph=KnowledgeGraphConfig(enable_graph=False),
@@ -122,7 +122,7 @@ def smart_config():
         history_db_path=os.path.join(data_dir, "history.db"),
         collection_name="dhee_memories",
         embedding_model_dims=dims,
-        engram=FadeMemConfig(enable_forgetting=True),
+        fade=FadeMemConfig(enable_forgetting=True),
         echo=EchoMemConfig(enable_echo=has_llm),
         category=CategoryMemConfig(enable_categories=has_llm),
         graph=KnowledgeGraphConfig(enable_graph=True, use_llm_extraction=False),

diff --git a/dhee/core/forgetting.py b/dhee/core/forgetting.py
@@ -29,14 +29,14 @@ def __init__(
         self,
         db: "SQLiteManager",
         config: "DistillationConfig",
-        fadem_config: "FadeMemConfig",
+        fade_config: "FadeMemConfig",
         resolve_conflict_fn=None,
         search_fn=None,
         llm=None,
     ):
         self.db = db
         self.config = config
-        self.fadem_config = fadem_config
+        self.fade_config = fade_config
         self.resolve_conflict_fn = resolve_conflict_fn
         self.search_fn = search_fn
         self.llm = llm
@@ -90,7 +90,7 @@ def run(
                 nearest = neighbors[0]
                 similarity = float(nearest.score)
 
-                if similarity < self.fadem_config.conflict_similarity_threshold:
+                if similarity < self.fade_config.conflict_similarity_threshold:
                     continue
 
                 # Fetch the neighbor memory from DB
@@ -319,12 +319,12 @@ def __init__(
         self,
         db: "SQLiteManager",
         config: "DistillationConfig",
-        fadem_config: "FadeMemConfig",
+        fade_config: "FadeMemConfig",
         delete_fn=None,
     ):
         self.db = db
         self.config = config
-        self.fadem_config = fadem_config
+        self.fade_config = fade_config
         self.delete_fn = delete_fn
 
     def run(
@@ -373,7 +373,7 @@ def run(
                 pressure = strength * pressure_factor * excess_ratio
                 new_strength = max(0.0, strength - pressure)
 
-                if new_strength < self.fadem_config.forgetting_threshold:
+                if new_strength < self.fade_config.forgetting_threshold:
                     if self.delete_fn:
                         try:
                             self.delete_fn(memory["id"])

diff --git a/dhee/mcp_server.py b/dhee/mcp_server.py
@@ -57,7 +57,7 @@ def _get_embedding_dims_for_model(model: str, provider: str) -> int:
         "text-embedding-3-large": 3072,
         "text-embedding-ada-002": 1536,
     }
-    env_dims = os.environ.get("FADEM_EMBEDDING_DIMS")
+    env_dims = os.environ.get("DHEE_EMBEDDING_DIMS") or os.environ.get("FADEM_EMBEDDING_DIMS")
     if env_dims:
         return int(env_dims)
     if model in EMBEDDING_DIMS:
@@ -69,8 +69,8 @@ def _get_embedding_dims_for_model(model: str, provider: str) -> int:
     return 3072
 
 
-def get_memory_instance() -> Memory:
-    """Create and return a configured Memory instance (FullMemory for MCP)."""
+def get_memory_instance() -> FullMemory:
+    """Create and return a configured FullMemory instance for the MCP server."""
     openai_key = os.environ.get("OPENAI_API_KEY")
     gemini_key = os.environ.get("GOOGLE_API_KEY") or os.environ.get("GEMINI_API_KEY")
     nvidia_key = (
@@ -79,13 +79,17 @@ def get_memory_instance() -> Memory:
         or os.environ.get("NVIDIA_EMBEDDING_API_KEY")
     )
 
+    def _env(key: str, default: str = "") -> str:
+        """Read DHEE_ env var with FADEM_ fallback for backward compat."""
+        return os.environ.get(f"DHEE_{key}") or os.environ.get(f"FADEM_{key}") or default
+
     if openai_key:
-        embedder_model = os.environ.get("FADEM_EMBEDDER_MODEL", "text-embedding-3-small")
+        embedder_model = _env("EMBEDDER_MODEL", "text-embedding-3-small")
         embedding_dims = _get_embedding_dims_for_model(embedder_model, "openai")
         llm_config = LLMConfig(
             provider="openai",
             config={
-                "model": os.environ.get("FADEM_LLM_MODEL", "gpt-4o-mini"),
+                "model": _env("LLM_MODEL", "gpt-4o-mini"),
                 "temperature": 0.1, "max_tokens": 1024, "api_key": openai_key,
             }
         )
@@ -94,12 +98,12 @@ def get_memory_instance() -> Memory:
             config={"model": embedder_model, "api_key": openai_key},
         )
     elif gemini_key:
-        embedder_model = os.environ.get("FADEM_EMBEDDER_MODEL", "gemini-embedding-001")
+        embedder_model = _env("EMBEDDER_MODEL", "gemini-embedding-001")
         embedding_dims = _get_embedding_dims_for_model(embedder_model, "gemini")
         llm_config = LLMConfig(
             provider="gemini",
             config={
-                "model": os.environ.get("FADEM_LLM_MODEL", "gemini-2.0-flash"),
+                "model": _env("LLM_MODEL", "gemini-2.0-flash"),
                 "temperature": 0.1, "max_tokens": 1024, "api_key": gemini_key,
             }
         )
@@ -108,13 +112,12 @@ def get_memory_instance() -> Memory:
             config={"model": embedder_model, "api_key": gemini_key},
         )
     elif nvidia_key:
-        # Internal provider — not customer-documented
-        embedder_model = os.environ.get("FADEM_EMBEDDER_MODEL", "nvidia/llama-nemotron-embed-vl-1b-v2")
+        embedder_model = _env("EMBEDDER_MODEL", "nvidia/llama-nemotron-embed-vl-1b-v2")
         embedding_dims = 2048
         llm_config = LLMConfig(
             provider="nvidia",
             config={
-                "model": os.environ.get("FADEM_LLM_MODEL", "qwen/qwen3.5-397b-a17b"),
+                "model": _env("LLM_MODEL", "qwen/qwen3.5-397b-a17b"),
                 "temperature": 0.2, "max_tokens": 4096, "api_key": nvidia_key,
             }
         )
@@ -131,17 +134,15 @@ def get_memory_instance() -> Memory:
         )
 
     from dhee.configs.base import _dhee_data_dir
-    vec_db_path = os.environ.get(
-        "FADEM_VEC_DB_PATH",
-        os.path.join(_dhee_data_dir(), "zvec"),
-    )
+    vec_db_path = _env("VEC_DB_PATH") or os.path.join(_dhee_data_dir(), "zvec")
+    collection = _env("COLLECTION", "dhee_memories")
 
     # Use in-memory vector store for simple embedder (no persistent storage needed)
     if embedder_config.provider == "simple":
         vector_store_config = VectorStoreConfig(
             provider="memory",
             config={
-                "collection_name": os.environ.get("FADEM_COLLECTION", "fadem_memories"),
+                "collection_name": collection,
                 "embedding_model_dims": embedding_dims,
             },
         )
@@ -150,20 +151,17 @@ def get_memory_instance() -> Memory:
             provider="zvec",
             config={
                 "path": vec_db_path,
-                "collection_name": os.environ.get("FADEM_COLLECTION", "fadem_memories"),
+                "collection_name": collection,
                 "embedding_model_dims": embedding_dims,
             },
         )
 
-    history_db_path = os.environ.get(
-        "FADEM_HISTORY_DB",
-        os.path.join(_dhee_data_dir(), "history.db"),
-    )
+    history_db_path = _env("HISTORY_DB") or os.path.join(_dhee_data_dir(), "history.db")
 
-    fadem_config = FadeMemConfig(
-        enable_forgetting=os.environ.get("FADEM_ENABLE_FORGETTING", "true").lower() == "true",
-        sml_decay_rate=float(os.environ.get("FADEM_SML_DECAY_RATE", "0.15")),
-        lml_decay_rate=float(os.environ.get("FADEM_LML_DECAY_RATE", "0.02")),
+    fade_config = FadeMemConfig(
+        enable_forgetting=_env("ENABLE_FORGETTING", "true").lower() == "true",
+        sml_decay_rate=float(_env("SML_DECAY_RATE", "0.15")),
+        lml_decay_rate=float(_env("LML_DECAY_RATE", "0.02")),
     )
 
     config = MemoryConfig(
@@ -172,18 +170,18 @@ def get_memory_instance() -> Memory:
         embedder=embedder_config,
         history_db_path=history_db_path,
         embedding_model_dims=embedding_dims,
-        engram=fadem_config,
+        fade=fade_config,
     )
 
     return FullMemory(config)
 
 
 # Global instances (lazy)
-_memory: Optional[Memory] = None
+_memory: Optional[FullMemory] = None
 _buddhi = None  # type: ignore
 
 
-def get_memory() -> Memory:
+def get_memory() -> FullMemory:
     global _memory
     if _memory is None:
         _memory = get_memory_instance()
@@ -203,7 +201,7 @@ def get_buddhi():
 
 server = Server("dhee")
 
-# Tool definitions — 8 tools total
+# Tool definitions — 24 tools
 TOOLS = [
     Tool(
         name="remember",

diff --git a/dhee/memory/core.py b/dhee/memory/core.py
@@ -65,7 +65,7 @@ def __init__(
         self.vector_store = VectorStoreFactory.create(
             self.config.vector_store.provider, self.config.vector_store.config
         )
-        self.fadem_config = self.config.engram
+        self.fade_config = self.config.fade
         self.distillation_config = getattr(self.config, "distillation", None)
 
         # Query embedding LRU cache
@@ -115,7 +115,7 @@ def add(
             # Boost fast trace if multi-trace is enabled
             if self.distillation_config and self.distillation_config.enable_multi_trace:
                 s_fast = existing.get("s_fast") or 0.0
-                boosted = boost_fast_trace(s_fast, self.fadem_config.access_strength_boost)
+                boosted = boost_fast_trace(s_fast, self.fade_config.access_strength_boost)
                 self.db.update_memory(existing["id"], {"s_fast": boosted})
             return {
                 "results": [{
@@ -162,7 +162,7 @@ def add(
             "confidentiality_scope": metadata.get("confidentiality_scope", "work"),
             "source_type": "mcp",
             "source_app": source_app,
-            "decay_lambda": self.fadem_config.sml_decay_rate,
+            "decay_lambda": self.fade_config.sml_decay_rate,
             "status": "active",
             "importance": metadata.get("importance", 0.5),
             "sensitivity": metadata.get("sensitivity", "normal"),
@@ -374,11 +374,11 @@ def apply_decay(
                 last_accessed=mem.get("last_accessed", mem.get("created_at", "")),
                 access_count=int(mem.get("access_count", 0)),
                 layer=mem.get("layer", "sml"),
-                config=self.fadem_config,
+                config=self.fade_config,
             )
 
-            if should_forget(new_strength, self.fadem_config):
-                if self.fadem_config.use_tombstone_deletion:
+            if should_forget(new_strength, self.fade_config):
+                if self.fade_config.use_tombstone_deletion:
                     self.db.update_memory(mem["id"], {"tombstone": 1, "strength": new_strength})
                 else:
                     self.db.delete_memory(mem["id"])
@@ -391,7 +391,7 @@ def apply_decay(
                 mem.get("layer", "sml"),
                 int(mem.get("access_count", 0)),
                 new_strength,
-                self.fadem_config,
+                self.fade_config,
             ):
                 self.db.update_memory(mem["id"], {"strength": new_strength, "layer": "lml"})
                 promoted += 1