feat: integrate memory spaces and rolling review

Author: AnkushMalaker

.gitignore

@@ -87,6 +87,17 @@ extras/speaker-omni-experimental/cache/*
 # Discovery agent PID file
 edge/.discovery-agent.pid
 
+# Runtime-generated diagnostics and reports. Keep durable documentation under docs/;
+# these locations are regenerated by local runs and screenshot/evaluation tooling.
+/artifacts/
+/backends/advanced/artifacts/
+
+# Runtime and recovered Chronicle memory vaults. The canonical backend vault is
+# already covered by backends/advanced/data/* above; the recursive rule also keeps
+# copied vault trees from accidentally becoming Git-visible elsewhere.
+**/conversation_docs/
+/recovered-vault/
+
 # Local Graphify knowledge graph. Agents query this directly; it is generated
 # per checkout and hidden from the editor explorer via .vscode/settings.json.
 graphify-out/

CONTEXT.md

@@ -30,6 +30,14 @@ objects. Capture and meaning are different layers.
   audio ranges are authoritative; related Conversation IDs are lineage/navigation.
 - **Vault**: the Markdown memory source of truth. It is derived from settled semantic
   inputs and can be regenerated from durable database evidence.
+- **Potential memory**: a proposed change derived from one reviewed day and the
+  currently accepted Vault. It is not memory and is invisible to retrieval until a
+  person accepts it.
+- **Review day**: one local day whose Episode account and Potential memories are being
+  decided. Review days advance in chronological order.
+- **Unexplained evidence**: captured evidence inside a reviewed interval that no active
+  Timeline episode accounts for. It is distinct from a capture gap, where no evidence
+  exists to interpret.
 
 ## Invariants
 
@@ -70,6 +78,9 @@ objects. Capture and meaning are different layers.
     conversation. Global filters, enrollment, memory, and corpus analysis must collapse
     unknowns into a category or key them by conversation plus local label; they must
     never equate matching placeholder text across conversations.
+17. A later Review day never observes an unresolved Potential memory. It is derived
+    only after every earlier Review day has either changed the accepted Vault or been
+    rejected without changing it.
 
 ## Important Modules and Interfaces
 

backends/advanced/docker-compose.yml

@@ -220,22 +220,24 @@ services:
   # for HTTPS. compose-stack.md#webui-dev
   webui-dev:
     build:
-      context: ./webui
-      dockerfile: Dockerfile.dev
+      context: ../..
+      dockerfile: backends/advanced/webui/Dockerfile.dev
     ports:
       - "${WEBUI_DEV_PORT:-5173}:5173"
     environment:
       - VITE_BACKEND_URL=${VITE_BACKEND_URL-http://${HOST_IP}:${BACKEND_PUBLIC_PORT:-8000}}
       - VITE_ALLOWED_HOSTS=${VITE_ALLOWED_HOSTS:-localhost 127.0.0.1}
       - VITE_HMR_PORT=${VITE_HMR_PORT:-5173}
     volumes:
-      - ./webui/src:/app/src
-      - ./webui/public:/app/public
+      - ./webui/src:/workspace/backends/advanced/webui/src
+      - ../../contracts/audio/v2/typescript:/workspace/contracts/audio/v2/typescript:ro
+      - ./webui/public:/workspace/backends/advanced/webui/public
       # Vite's entry document. Baked into the image otherwise, so edits to the
       # page title/meta silently kept serving the built-in copy until a rebuild.
-      - ./webui/index.html:/app/index.html:ro
-      - ./webui/tailwind.config.js:/app/tailwind.config.js:ro
-      - ./webui/chronicle-espresso-preset.js:/app/chronicle-espresso-preset.js:ro
+      - ./webui/index.html:/workspace/backends/advanced/webui/index.html:ro
+      - ./webui/vite.config.ts:/workspace/backends/advanced/webui/vite.config.ts:ro
+      - ./webui/tailwind.config.js:/workspace/backends/advanced/webui/tailwind.config.js:ro
+      - ./webui/chronicle-espresso-preset.js:/workspace/backends/advanced/webui/chronicle-espresso-preset.js:ro
     depends_on:
       chronicle-backend:
         condition: service_healthy
@@ -287,6 +289,7 @@ services:
     volumes:
       - ./data/syncthing_config:/var/syncthing/config
       - ./data/conversation_docs:/vaults   # backend exposes per-user folders under here
+      - ./data/memory_spaces:/memory-spaces # isolated space vaults are sibling folders
     ports:
       - "22000:22000/tcp"   # sync protocol (reachable via Tailscale)
       - "22000:22000/udp"   # QUIC

backends/advanced/init.py

@@ -48,9 +48,6 @@
 
 # Match the Pi executor's conservative fallbacks without baking one machine's
 # served context into every model selected by the wizard.
-DEFAULT_PI_CONTEXT_WINDOW = 32768
-DEFAULT_PI_MAX_TOKENS = 4096
-PI_PROMPT_HEADROOM_TOKENS = 1024
 MANAGED_LLAMACPP_REGISTRY_MODELS = {"llamacpp-llm", "muse-glimmer-llm"}
 
 
@@ -1349,10 +1346,6 @@ def setup_memory_agents(self):
         if not isinstance(existing_pi, dict):
             existing_pi = {}
 
-        configured_llm = str(
-            self.config_manager.get_config_defaults().get("llm") or "muse-glimmer-llm"
-        )
-
         codex_config = {
             "model": "gpt-5.6-luna",
             "reasoning_effort": "low",
@@ -1363,60 +1356,51 @@ def setup_memory_agents(self):
             **existing_codex,
         }
         pi_config = {
-            "model": configured_llm,
             "timeout_seconds": 900,
             "thinking": "off",
             **existing_pi,
         }
-        if not str(pi_config.get("model") or "").strip():
-            pi_config["model"] = configured_llm
+        # Model routing is owned by defaults + llm_operations. Older wizard runs
+        # copied that policy into the executor adapter, causing provider switches to
+        # require a second unrelated edit.
+        pi_config.pop("model", None)
 
         if "pi" in (selected_write, selected_search, recovery_backend):
             effective_models = self._effective_model_registry()
-            llm_models = sorted(
-                name
-                for name, model in effective_models.items()
-                if str(model.get("model_type") or "").lower() == "llm"
-                and str(model.get("api_family") or "").lower() == "openai"
-            )
-            if llm_models:
-                self.console.print(
-                    "[blue][INFO][/blue] Available OpenAI-compatible LLM registry entries: "
-                    + ", ".join(llm_models)
-                )
-            pi_model = self.prompt_value(
-                "Chronicle model registry entry for Pi",
-                str(pi_config.get("model") or configured_llm),
-            ).strip()
-            if not pi_model:
-                raise ValueError("Pi requires a Chronicle model registry entry")
-            selected_model = effective_models.get(pi_model)
-            if selected_model is None:
-                raise ValueError(
-                    f"Pi model registry entry '{pi_model}' does not exist in the "
-                    "effective defaults.yml + config.yml registry"
-                )
-            if str(selected_model.get("model_type") or "").lower() != "llm":
-                raise ValueError(f"Pi model registry entry '{pi_model}' is not an LLM")
-            if str(selected_model.get("api_family") or "").lower() != "openai":
-                raise ValueError(
-                    f"Pi model registry entry '{pi_model}' is not OpenAI-compatible"
-                )
-
-            previous_model = str(existing_pi.get("model") or "").strip()
-            model_changed = bool(previous_model and previous_model != pi_model)
-            if "context_window" not in existing_pi or model_changed:
-                context_window = self._pi_context_window(selected_model)
-                pi_config["context_window"] = context_window
-            else:
-                context_window = int(pi_config["context_window"])
-            if "max_tokens" not in existing_pi or model_changed:
-                pi_config["max_tokens"] = self._pi_max_tokens(context_window)
-            pi_config["model"] = pi_model
+            full_config = self.config_manager.get_full_config()
+            defaults = full_config.get("defaults") or {}
+            operations = full_config.get("llm_operations") or {}
+            operation_names = set()
+            if "pi" in (selected_write, recovery_backend):
+                operation_names.add("memory_write")
+            if selected_search == "pi":
+                operation_names.add("memory_search")
+            resolved_names = []
+            for operation_name in sorted(operation_names):
+                operation_config = operations.get(operation_name) or {}
+                model_name = str(
+                    operation_config.get("model") or defaults.get("llm") or ""
+                ).strip()
+                selected_model = effective_models.get(model_name)
+                if selected_model is None:
+                    raise ValueError(
+                        f"{operation_name} resolves to model registry entry "
+                        f"{model_name!r}, which does not exist"
+                    )
+                if str(selected_model.get("model_type") or "").lower() != "llm":
+                    raise ValueError(
+                        f"{operation_name} resolves to {model_name!r}, which is not an LLM"
+                    )
+                if str(selected_model.get("api_family") or "").lower() != "openai":
+                    raise ValueError(
+                        f"{operation_name} resolves to {model_name!r}, which is not "
+                        "OpenAI-compatible"
+                    )
+                resolved_names.append(f"{operation_name} → {model_name}")
             self.console.print(
-                "[green][SUCCESS][/green] Pi will use model registry entry: "
-                f"{pi_model}. No Pi auth directory or host login is required; the "
-                "backend resolves the model URL/key and creates isolated runtime config."
+                "[green][SUCCESS][/green] Pi will use central model routing: "
+                + ", ".join(resolved_names)
+                + ". No Pi auth directory or host login is required."
             )
 
         if "codex" in (selected_write, recovery_backend):
@@ -1508,30 +1492,6 @@ def _effective_model_registry(self) -> Dict[str, Dict[str, Any]]:
                     effective[str(model["name"])] = dict(model)
         return effective
 
-    @staticmethod
-    def _pi_context_window(model: Dict[str, Any]) -> int:
-        """Use a model-declared context, falling back to Pi's safe generic window."""
-        model_params = model.get("model_params") or {}
-        raw = model.get("context_window")
-        if raw in (None, "") and isinstance(model_params, dict):
-            raw = model_params.get("context_window")
-        try:
-            context_window = int(raw)
-        except (TypeError, ValueError):
-            return DEFAULT_PI_CONTEXT_WINDOW
-        if context_window <= PI_PROMPT_HEADROOM_TOKENS:
-            return DEFAULT_PI_CONTEXT_WINDOW
-        return context_window
-
-    @staticmethod
-    def _pi_max_tokens(context_window: int) -> int:
-        """Keep most context for prompts/tools while allowing larger-context models more output."""
-        return min(
-            DEFAULT_PI_MAX_TOKENS,
-            context_window // 4,
-            context_window - PI_PROMPT_HEADROOM_TOKENS,
-        )
-
     def setup_optional_services(self):
         """Configure optional services"""
         # Check if speaker service URL provided via args

backends/advanced/src/advanced_omi_backend/app_factory.py

@@ -51,6 +51,11 @@
 )
 from advanced_omi_backend.models.manual_memory import ManualMemory
 from advanced_omi_backend.models.memory_audit import MemoryAuditEntry
+from advanced_omi_backend.models.memory_space import (
+    DeferredSpaceEvent,
+    MemorySpace,
+    SpaceMergeProposal,
+)
 from advanced_omi_backend.models.system_event import SystemEvent
 from advanced_omi_backend.models.timeline import (
     AudioEvidenceSpan,
@@ -170,6 +175,9 @@ async def lifespan(app: FastAPI):
                 Annotation,
                 MemoryAuditEntry,
                 ManualMemory,
+                MemorySpace,
+                SpaceMergeProposal,
+                DeferredSpaceEvent,
                 SystemEvent,
                 CaptureSource,
                 PairingCode,