feat(memory): rebuild durable learning and management

AGENTS.md

@@ -173,7 +173,7 @@ packages/agent-core/src/
 ├── agents/tool-filter.test.ts  # Architecture coverage for definition-based capability filtering
 ├── agents/query/               # runLlmStream + tool execution cycle (max 50 steps), doom detection
 ├── agents/query/loop-hooks.ts  # 4 hook points: beforeModelBuild, beforeModelCall, afterStepEnd, afterLoopEnd
-├── agents/query/hooks/         # auto-compact, auto-inject-reminder, title-generation, todo-continuation, memory-extraction, memory-consolidation
+├── agents/query/hooks/         # auto-compact, auto-inject-reminder, title-generation, todo-continuation
 ├── execution/session-execution-manager.ts # Sole logical Execution lifecycle/admission, run resources, abort, recovery, and terminal owner
 ├── process/                    # ProcessRunner lifecycle, bounded streaming, timeout/abort, and structured results
 ├── tools/define-tool.ts        # defineTool() → ToolDescriptor (strict RawToolResult + explicit outputPolicy)
@@ -188,11 +188,11 @@ packages/agent-core/src/
 ├── tools/riipgrep/             # Ripgrep wrapper for search tools
 ├── core/                       # register-tools.ts: wires tools and finalized-result audit/logger hooks
 ├── store/                      # Zustand vanilla store: createSessionStore, StreamEvent reducer, ModelMessage projection, persist/load
-├── background/                 # BackgroundTaskManager (fire-and-forget, dedup) + tasks: title-generation, memory-extraction, memory-consolidation
+├── background/                 # BackgroundTaskManager (fire-and-forget, dedup) + title-generation task
 ├── commands/                   # CommandRegistry + /compact command
 ├── compression/                # DCP-like dynamic range compression: model tool action, refs, block state, soft/strong nudges below hard threshold
 ├── compact/                    # Mandatory hard compact safety path at >=85% context pressure plus /compact command
-├── memory/                     # MemoryFileManager (atomic writes, frontmatter, index), schemas, types, constants
+├── memory/                     # MemoryService, Markdown adapter, idle learning coordinator, policy, schemas, limits
 ├── session-goal/               # Session.goal schema, ownership service, status, budget, and usage
 ├── hitl/                       # Durable project-scoped approval/question queue and redacted display payloads
 ├── automations/                # Canonical Automation schemas, schedule, durable Invocation, Session dispatch
@@ -420,9 +420,13 @@ All six implement `Agent`: `store: StoreApi<SessionStoreState>`, `run(options) 
 beforeModelBuild (auto-compact) → toModelMessages → beforeModelCall (auto-inject-reminder)
   → runLlmStream → consumeFullStream → afterStepEnd (todo-continuation)
   → executeToolCalls (doom detection → partition → guards → execute)
-→ afterLoopEnd (todo-continuation, memory-extraction, memory-consolidation)
+→ afterLoopEnd (todo-continuation)
 ```
 
+Successful root Lead/Discussion terminals update the durable Memory cursor;
+`MemoryIdleCoordinator` performs automatic learning outside the Query Loop after
+10 minutes of inactivity.
+
 ## Tool System
 
 **35+ builtin tools** (base tools via `createBuiltinToolDescriptors()`, memory, Goal, Automation, Project Todo, and GitHub connector tools — all registered in `core/register-tools.ts`):
@@ -457,7 +461,7 @@ ArchCode has two intentionally separate context-reduction paths. Dynamic DCP-lik
 
 ## Memory System
 
-Project: `.archcode/runtime/memory/`, User: `~/.archcode/memory/` (user-global, not under project runtime). Structure: `index.md` (topic index), `preferences.md`, `knowledge/{topic}.md` (frontmatter + markdown). Types: `"user" | "feedback" | "project" | "reference"`. `MemoryFileManager`: atomic writes, path validation, frontmatter parse/format, index rebuild/search. Extraction (background task via `runLlmObject`) → writes topics. Consolidation (background task) → reorganizes index. Injection: ConfiguredAgent resolves one immutable Execution snapshot; PromptContractCompiler labels it non-authoritative and emits its source/status in the durable Prompt trace. `memory_write` rejects secrets.
+Project: `.archcode/runtime/memory/`, User: `~/.archcode/memory/` (user-global, not under project runtime). Structure: `index.md` (generated topic index), `preferences.md`, `knowledge/{topic}.md` (frontmatter + Markdown). Types: `"user" | "feedback" | "project" | "reference"`. `MemoryService` is the sole mutation boundary over `MemoryFileManager`: it owns CAS revisions, secret rejection, 8 KiB preferences, 16 KiB complete topic documents, the 200-topic cap, legacy shrink-only edits, index rebuilds, and deterministic receipt replay. Existing `memory_write` remains the immediate explicit-write path. Automatic learning is owned by the runtime-scoped `MemoryIdleCoordinator`: successful root Lead/Discussion conversations wait for 10 minutes of inactivity, then use at most one `fast` extraction call and one full-file reconciliation call; durable cursors, receipts, policy epochs, and warnings make restart and opt-out behavior explicit. Injection: ConfiguredAgent resolves one immutable Execution snapshot containing complete in-capacity preferences and project index; PromptContractCompiler labels it non-authoritative and emits its source/status in the durable Prompt trace.
 
 ## Session Goal System
 

README.md

@@ -117,6 +117,15 @@ Use an Automation when work should start once or on a recurring schedule. Goals
 and Automations are optional; both remain visible through Sessions inside the
 same workbench.
 
+## Durable, inspectable memory
+
+ArchCode keeps personal preferences and project knowledge as ordinary Markdown.
+Explicit requests to remember something are saved immediately through the
+Memory tool. Other durable context is considered only after a successful root
+conversation has been idle for 10 minutes, then reconciled against the complete
+Memory files it actually affects. Settings → Memory lets you inspect, edit,
+delete, disable recall, or opt out of automatic learning without deleting data.
+
 ## Run it your way
 
 | Where ArchCode runs | Good for | How you open it |

apps/server/src/app.test.ts

@@ -29,6 +29,10 @@ describe("createRuntimeApp", () => {
       subscribeSessionEvents: mock((listener: (event: GlobalSSEEvent) => void) => {
         listener({ type: "event", slug: "proj", sessionId: "session-1", eventId: 1, createdAt: 1, agentName: "lead", payload: {
           type: "execution-start",
+          memoryPolicy: {
+            policy: { useMemory: true, autoLearning: true },
+            epoch: { bootId: "test-memory-boot", generation: 0 },
+          },
           executionId: "run-1",
           binding: {
             selection: { model: "local:test" },

apps/server/src/app.ts

@@ -18,6 +18,7 @@ import { createAutomationsRoutes } from "./routes/automations";
 import { createAttachmentsRoutes } from "./routes/attachments";
 import { createMessagesRoutes } from "./routes/messages";
 import { createMcpRoutes } from "./routes/mcp";
+import { createMemoryRoutes } from "./routes/memory";
 import { createProjectsRoutes } from "./routes/projects";
 import { createSessionsRoutes } from "./routes/sessions";
 import { createTodosRoutes } from "./routes/todos";
@@ -82,6 +83,7 @@ export function createRuntimeApp(
   const files = createFilesRoutes(serverRuntime);
   const directories = createDirectoriesRoutes();
   const mcp = createMcpRoutes(serverRuntime);
+  const memory = createMemoryRoutes(serverRuntime);
 
   app.route("/api", globalWork);
   app.route("/api/projects", projects);
@@ -95,6 +97,7 @@ export function createRuntimeApp(
   app.route("/api/projects/:slug/sessions/:sessionId/tool-outputs", toolOutputs);
   app.route("/api/events", globalEvents);
   app.route("/api/projects", files);
+  app.route("/api/projects", memory);
   app.route("/api/mcp", mcp);
   app.route("/api/sessions", new Hono());
   app.route("/api/agents", agents);

apps/server/src/errors.ts

@@ -28,6 +28,12 @@ export type ServerErrorCode =
   | "CONFIG_REVISION_CONFLICT"
   | "CONFIG_VALIDATION_ERROR"
   | "CONFIG_RECOVERY_CONFLICT"
+  | "MEMORY_NOT_FOUND"
+  | "MEMORY_REVISION_CONFLICT"
+  | "MEMORY_CAPACITY_EXCEEDED"
+  | "MEMORY_INVALID_INPUT"
+  | "MEMORY_SECRET_DETECTED"
+  | "MEMORY_OPERATION_FAILED"
   | "TOOL_OUTPUT_FORBIDDEN"
   | "TOOL_OUTPUT_NOT_FOUND"
   | "TOOL_OUTPUT_EXPIRED"

apps/server/src/routes/config-recovery.ts

@@ -263,7 +263,7 @@ function safeIssuePath(path: string, configPath: string): string {
   }
   if (segments[0] === "memory") {
     const field = segments[1];
-    return new Set(["enabled", "minMessages", "minContentLength", "cooldownMs"]).has(field ?? "")
+    return new Set(["useMemory", "autoLearning"]).has(field ?? "")
       ? `memory.${field}`
       : "memory.<field>";
   }