Merge pull request #9 from Sankhya-AI/alpha

Release-V1

Author: Ashish-dwi99

README.md

@@ -133,7 +133,8 @@ You have three agents configured: Claude Code for deep reasoning, Codex for fast
 | **No episodic memory** | Vector search only | CAST scenes — time/place/topic clustering |
 | **No consolidation** | Store everything as-is | CLS sleep cycles — episodic to semantic distillation |
 | **No real-time coordination** | Polling or nothing | Active memory signal bus — agents see each other instantly |
-| **Concurrent access** | Single-process locks | sqlite-vec WAL — multiple agents, one DB |
+| **Agents don't learn** | Retrain or nothing | Skill-policy memory — agents accumulate reusable skills |
+| **Concurrent access** | Single-process locks | zvec HNSW — multiple agents, directory-based collections |
 
 ---
 
@@ -155,6 +156,35 @@ Agent capabilities are stored as memories: *"claude-code: Advanced coding agent.
 
 No new database tables. No separate routing service. The same `Memory.add()` / `Memory.search()` that stores user conversations also stores agent profiles and routes tasks.
 
+### Skill Memory — the self-improvement loop
+
+Agents learn from experience. When an agent completes a task, Engram records the trajectory (actions, tools, results). Successful trajectories accumulate. The Skill Miner analyzes clusters of similar trajectories and extracts reusable **skills** — validated procedures stored as SKILL.md files with YAML frontmatter.
+
+Skills have confidence scores that update on success/failure (Bayesian, asymmetric — failures penalize more). High-confidence skills are automatically suggested when matching tasks arrive. The loop:
+
+```
+Agent works → Trajectory recorded → Miner extracts patterns → Skills stored
+  ↑                                                               |
+  └── Agent applies skill → Outcome logged → Confidence updated ──┘
+```
+
+```python
+from engram import SmartMemory
+
+m = SmartMemory(preset="smart")
+
+# Search for relevant skills
+skills = m.search_skills("fix python import error")
+
+# Apply a skill — returns injectable recipe
+result = m.apply_skill(skill_id)
+
+# Report outcome — updates confidence
+m.log_skill_outcome(skill_id, success=True)
+```
+
+Skills are discovered from `~/.engram/skills/` and `{repo}/.engram/skills/`. Six MCP tools: `search_skills`, `apply_skill`, `log_skill_outcome`, `record_trajectory_step`, `mine_skills`, `get_skill_stats`.
+
 ### Handoff
 
 When an agent pauses (rate limit, crash, tool switch), it saves a session digest: task summary, decisions made, files touched, TODOs remaining. The next agent loads it and continues. If no digest was saved, Engram falls back to parsing the conversation logs automatically.
@@ -171,6 +201,8 @@ When an agent pauses (rate limit, crash, tool switch), it saves a session digest
 | **CLS Distillation** | Sleep-cycle replay: episodic to semantic fact extraction |
 | **Multi-trace** | Benna-Fusi model — fast/mid/slow decay traces per memory |
 | **Intent routing** | Episodic vs semantic query classification |
+| **Skill Memory** | SKILL.md files — discover, apply, and mine reusable agent skills |
+| **Skill Miner** | Trajectory recording → pattern extraction → skill compilation |
 | **Orchestrator** | Agent registry + semantic task routing + CAS claim/release |
 | **Handoff bus** | Session digests, checkpoints, JSONL log fallback |
 | **Active Memory** | Real-time signal bus with TTL tiers |
@@ -206,7 +238,7 @@ memory.add("User prefers Python over TypeScript", user_id="u1")
 results = memory.search("programming preferences", user_id="u1")
 ```
 
-**18 MCP tools** — memory CRUD, semantic search, episodic scenes, profiles, decay, session handoff. One command configures Claude Code, Cursor, and Codex:
+**14 MCP tools** — memory CRUD, semantic search, session handoff, skill search/apply/mine, trajectory recording. One command configures Claude Code, Cursor, and Codex:
 
 ```bash
 engram install
@@ -342,13 +374,14 @@ Works with any tool-calling agent via REST: `engram-api` starts a server at `htt
 ```
 ├── engram/                  # engram-memory — core Python package
 │   ├── core/                #   decay, echo, category, scenes, distillation, traces
-│   ├── memory/              #   Memory class (orchestrates all layers)
+│   ├── memory/              #   CoreMemory → SmartMemory → FullMemory
+│   ├── skills/              #   skill schema, store, discovery, executor, miner, trajectories
 │   ├── llms/                #   LLM providers (gemini, openai, nvidia, ollama)
 │   ├── embeddings/          #   embedding providers
-│   ├── vector_stores/       #   sqlite-vec, in-memory
+│   ├── vector_stores/       #   zvec, sqlite-vec, in-memory
 │   ├── db/                  #   SQLite persistence
 │   ├── api/                 #   REST API endpoints
-│   ├── mcp_server.py        #   MCP server (18 tools)
+│   ├── mcp_server.py        #   MCP server (14 tools)
 │   └── cli.py               #   CLI interface
 ├── engram-bus/              # engram-bus — agent communication
 │   └── engram_bus/          #   bus, pub/sub, handoff store, TCP server

engram-bus/tests/__init__.py

@@ -14,7 +14,8 @@
 from fastapi.staticfiles import StaticFiles
 from pydantic import BaseModel, Field
 
-from engram import Memory
+from engram.configs.base import EmbedderConfig, LLMConfig, MemoryConfig, VectorStoreConfig
+from engram.memory.main import Memory
 from engram_enterprise.api.auth import (
     enforce_session_issuer,
     get_token_from_request,
@@ -41,6 +42,7 @@
     SessionCreateResponse,
 )
 from engram_enterprise.policy import feature_enabled
+from engram_enterprise.kernel import PersonalMemoryKernel
 from engram.exceptions import FadeMemValidationError
 from engram.observability import add_metrics_routes, logger as structured_logger, metrics
 
@@ -103,19 +105,54 @@ class DecayResponse(BaseModel):
 
 _memory: Optional[Memory] = None
 _memory_lock = threading.Lock()
+_kernel: Optional[PersonalMemoryKernel] = None
+_kernel_lock = threading.Lock()
+
+
+def _fallback_memory_config() -> MemoryConfig:
+    data_dir = os.path.join(os.path.expanduser("~"), ".engram")
+    os.makedirs(data_dir, exist_ok=True)
+    dims = 384
+    return MemoryConfig(
+        llm=LLMConfig(provider="mock", config={}),
+        embedder=EmbedderConfig(provider="simple", config={"embedding_dims": dims}),
+        vector_store=VectorStoreConfig(
+            provider="memory",
+            config={
+                "collection_name": "engram_enterprise",
+                "embedding_model_dims": dims,
+            },
+        ),
+        history_db_path=os.path.join(data_dir, "enterprise_history.db"),
+        collection_name="engram_enterprise",
+        embedding_model_dims=dims,
+    )
 
 
 def get_memory() -> Memory:
     global _memory
     if _memory is None:
         with _memory_lock:
             if _memory is None:
-                _memory = Memory()
+                try:
+                    _memory = Memory()
+                except Exception as exc:
+                    logger.warning(
+                        "Failed to initialize default Memory config (%s). "
+                        "Falling back to mock/simple in-memory configuration.",
+                        exc,
+                    )
+                    _memory = Memory(config=_fallback_memory_config())
     return _memory
 
 
 def get_kernel():
-    return get_memory().kernel
+    global _kernel
+    if _kernel is None:
+        with _kernel_lock:
+            if _kernel is None:
+                _kernel = PersonalMemoryKernel(get_memory())
+    return _kernel
 
 
 def _extract_content(messages: Optional[Union[str, List[Dict[str, Any]]]], content: Optional[str]) -> str: