diff --git a/MEMORY.md b/MEMORY.md
index 54d3a51..0282628 100644
--- a/MEMORY.md
+++ b/MEMORY.md
@@ -43,11 +43,9 @@
 - Required sequence before any gateway intervention: (1) run `openclaw gateway status`; (2) report findings to Tao; (3) prefer non-disruptive diagnosis first; (4) only touch gateway lifecycle if Tao explicitly approves the exact command.
 
 ## 7) Model policy (current preference)
-- Historical temporary preference once was: all subagents **primary** = `openai-codex/gpt-5.2`, **fallback** = `github-copilot/claude-opus-4.6`.
-- Current important exception / newer rule: **execution agents** (esp. `codex_worker`, and by default other workers unless Tao says otherwise) must target **primary** = `openai-codex/gpt-5.4`.
-  - **No silent fallback:** if the system falls back to any other model (or if `openai-codex/gpt-5.4` becomes unknown/unavailable), main must **immediately notify Tao**.
-  - **No auto-fallback for execution agents:** if 5.4 is unavailable, workers should not continue on another model; main should report and wait for Tao’s model decision.
-  - If an execution agent is not responding, main must consider **model unavailability/fallback** as a first-class suspected cause and **tell Tao**.
+- Current config (2026-03-21): all agents **primary** = `claude-local/claude-sonnet-4-6`, **fallback** = `openai-codex/gpt-5.4` (only one fallback; Minimax removed).
+- **铁律：Fallback 发生时必须通知 Tao。** 适用于**所有 agent**（main、tech_geek、travel_assistant、finance_assistant、所有 execution agents）。无论哪条链路触发了 fallback，都必须立即通知 Tao，说明：哪个 agent、从什么 model fallback 到了什么 model、原因（timeout / error / unavailable）。不允许静默 fallback。
+  - If an agent is not responding, main must consider **model unavailability/fallback** as a first-class suspected cause and **tell Tao**.
 - Additional long-lived execution agents initialized locally for repeat use: **docs_worker**, **qa_worker**, **ops_worker**.
 - Execution-agent system needs a standardized dispatch/handoff layer: task input template, result format, blocker/escalation rules, and a **main-to-Tao forwarding rule**.
 - **Hard reporting/forwarding protocol (must-follow):**
@@ -72,3 +70,9 @@
 - Local memory dir: `/Users/taodeng/.openclaw/workspace/main/memory/`.
 - Added `memory/INDEX.md` to classify daily notes vs topic notes vs automation state JSON (do not rename/move state JSON without updating jobs).
 - Plan: create a separate GitHub KB repo (option A) for shareable/curated knowledge; keep private/sensitive items local and only publish redacted content.
+
+## 11) Oracle Cloud
+- See [memory/oracle_cloud.md](memory/oracle_cloud.md) for SSH connection details and deployment notes.
+
+## 12) 开发原则 (Dev Principles)
+- See [memory/dev_principles.md](memory/dev_principles.md) for Tao's principles on new project development (when to reuse vs build, code quality, testing, versioning, etc.). Read before starting any new project.
diff --git a/memory/CURRENT_STATE.md b/memory/CURRENT_STATE.md
index 00aa6b5..17f4c47 100644
--- a/memory/CURRENT_STATE.md
+++ b/memory/CURRENT_STATE.md
@@ -1,33 +1,31 @@
 # Current State
-> Last updated: 2026-03-13T07:02:00+10:00
+> Last updated: 2026-03-21T02:00:58.027Z
 
 ## Objective
-Validate the memory-continuity lifecycle plugin path safely after upgrade.
+但是文件数目不可以超过500条
 
 ## Current Step
-Startup-injection testing is confirmed on multiple resident subagents after the upgrade; Discord main/channel/thread paths have now been tested and are not currently supported for reliable continuity recovery; compaction-path testing is now being prepared on `travel_assistant` as the disposable session.
+Conversation ended after 181 messages
 
 ## Key Decisions
-- Keep `SKILL.md + standard lifecycle plugin` as the primary architecture.
-- Treat ContextEngine as a future option, not the v1 default.
-- Do not force compaction on the main live session just to test Experiment C.
-- Use disposable resident subagent sessions for future compaction-path testing.
-- `travel_assistant` is now a valid sample again after removing the old wrong workspace and fixing its `AGENTS.md` startup rule to read `memory/CURRENT_STATE.md`.
-- Alpha support boundary is now explicit: resident subagent startup continuity is supported; Discord main/channel/thread continuity is not yet supported in `v0.3.0-probe`.
+- Auto-extracted from conversation
 
 ## Next Action
-Two-track next step:
-1. If continuing active testing, use `travel_assistant` as the sacrificial session, reload the updated probe logging, then force/accelerate a real compaction event and compare hook behavior against the failure chain seen on main.
-2. If pausing for upstream uncertainty, keep the fallback-continuity GitHub issue open and wait for OpenClaw feedback before spending more effort on the main-session fallback path.
+Continue from where we left off
 
 ## Blockers
-- Experiment C still lacks a real compaction event, so compaction-hook behavior remains unproven.
-- New upstream-suspected blocker: fallback continuity failure may be caused by OpenClaw providing incomplete plugin runtime (`workspaceDir=<missing>`) on fallback/new-session recovery path.
+None
 
 ## Unsurfaced Results
-- Experiment A is now confirmed on multiple resident subagents.
-- `tech_geek` recovered plugin-injected startup continuity state without using `read`.
-- `travel_assistant` also recovered the correct upgraded smoke-test state after fixing its old workspace pollution and startup rule.
-- Fresh Discord main/channel/thread tests failed to preserve short facts or concrete working-state recovery across new sessions; treat Discord main continuity as unsupported for this alpha.
-- New execution-model finding: main can successfully dispatch and consume results from temporary subagents (single, follow-up, and parallel two-worker tests all succeeded).
-- New ACP finding: Codex ACP worker is validated as a usable execution worker for main; Claude ACP today remains an invalid sample because Tao confirmed Claude was already in timeout/overuse state, so do not classify that failure as product/runtime evidence.
+🦞 大内总管: 已加硬上限。
+
+**改动：**
+- `MAX_MEMORY_FILES = 500`
+- 每次 archive 后检查 `memory/` 总文件数
+- 超过 500 → 自动删最旧的可删文件，降到 450（留 50 个余量）
+- **不会删的**：`CURRENT_STATE.md`、`MEMORY.md`、`INDEX.md`、主题文件、JSON 状态文件
+
+**三层防护：**
+1. 单类 archive 上限 20 个
+2. 总文件数硬上限 500
+3. 清理时按修改时间排序，从最旧的开始删
diff --git a/skills/memory-continuity/SKILL.md b/skills/memory-continuity/SKILL.md
index 57d0825..c12f273 100644
--- a/skills/memory-continuity/SKILL.md
+++ b/skills/memory-continuity/SKILL.md
@@ -130,6 +130,22 @@ Update the file by **overwriting** it, not appending, at these moments:
 | Before handoff / subagent exit | Preserves outputs and unsurfaced results |
 | After a substantive state change | Keeps checkpoint aligned with actual work |
 
+### Override rule
+
+**CURRENT_STATE.md must always be overwritten when:**
+- A new task or objective starts — regardless of what is currently in the file
+- The previous objective is complete or abandoned
+- The user gives a new task that supersedes the previous one
+
+Having content in CURRENT_STATE.md does NOT mean it should be preserved.
+Content only matters if Objective is still active and work is genuinely in progress.
+
+Checking before overwrite:
+- Read the file
+- If Objective matches the current task → update in place (overwrite)
+- If Objective does NOT match → overwrite the entire file with the new state
+- Never append. Never skip the update because "there's already something there".
+
 ### 4. Keep the checkpoint small
 
 `CURRENT_STATE.md` should usually stay under about 40 lines and be readable in
diff --git a/skills/memory-continuity/plugin/lifecycle-prototype.ts b/skills/memory-continuity/plugin/lifecycle-prototype.ts
index c4a3796..7f909ac 100644
--- a/skills/memory-continuity/plugin/lifecycle-prototype.ts
+++ b/skills/memory-continuity/plugin/lifecycle-prototype.ts
@@ -1,6 +1,18 @@
 import fs from "node:fs";
 import path from "node:path";
 
+const MAX_ARCHIVE_COUNT = 20;
+const MAX_MEMORY_FILES = 500;
+
+// Protected files that should never be auto-deleted
+const PROTECTED_FILES = new Set([
+  "CURRENT_STATE.md",
+  "MEMORY.md",
+  "INDEX.md",
+  "dev_principles.md",
+  "oracle_cloud.md",
+]);
+
 const PLACEHOLDER_VALUES = new Set([
   "",
   "none",
@@ -82,11 +94,70 @@ function formatArchiveStamp(date = new Date()) {
   ].join("-") + `_${pad(date.getHours())}-${pad(date.getMinutes())}`;
 }
 
+function pruneArchives(dir: string, pattern: RegExp, max: number) {
+  try {
+    const files = fs.readdirSync(dir)
+      .filter((f: string) => pattern.test(f))
+      .sort();
+    const excess = files.length - max;
+    if (excess > 0) {
+      for (let i = 0; i < excess; i++) {
+        fs.unlinkSync(path.join(dir, files[i]));
+      }
+      console.log(`[memory-continuity] pruned ${excess} old archive(s) in ${dir}`);
+    }
+  } catch {}
+}
+
+function enforceFileLimit(memoryDir: string) {
+  try {
+    const allFiles = fs.readdirSync(memoryDir, { recursive: true }) as string[];
+    // Only count files, not directories
+    const fileList = allFiles
+      .map((f: string) => ({ name: String(f), full: path.join(memoryDir, String(f)) }))
+      .filter((f) => {
+        try { return fs.statSync(f.full).isFile(); } catch { return false; }
+      });
+
+    if (fileList.length <= MAX_MEMORY_FILES) return;
+
+    // Sort deletable files by mtime (oldest first), skip protected
+    const deletable = fileList
+      .filter((f) => {
+        const base = path.basename(f.name);
+        if (PROTECTED_FILES.has(base)) return false;
+        if (base.endsWith(".json")) return false; // automation state
+        return true;
+      })
+      .map((f) => ({ ...f, mtime: fs.statSync(f.full).mtimeMs }))
+      .sort((a, b) => a.mtime - b.mtime);
+
+    const toDelete = fileList.length - 450; // bring down to 450 for headroom
+    const deleted = Math.min(toDelete, deletable.length);
+    for (let i = 0; i < deleted; i++) {
+      fs.unlinkSync(deletable[i].full);
+    }
+    if (deleted > 0) {
+      console.log(`[memory-continuity] enforced file limit: deleted ${deleted} oldest files (${fileList.length} → ${fileList.length - deleted})`);
+    }
+  } catch {}
+}
+
 function appendArchive(workspaceDir: string, markdown: string) {
   const stamp = formatArchiveStamp();
   const archiveDir = path.join(workspaceDir, "memory", "session_archive");
   fs.mkdirSync(archiveDir, { recursive: true });
   fs.writeFileSync(path.join(archiveDir, `${stamp}.md`), markdown, "utf8");
+
+  // Prune session_archive/ beyond MAX_ARCHIVE_COUNT
+  pruneArchives(archiveDir, /\.md$/, MAX_ARCHIVE_COUNT);
+
+  // Also prune STATE_ARCHIVE_*.md in memory/ root (generated by OpenClaw native)
+  const memoryDir = path.join(workspaceDir, "memory");
+  pruneArchives(memoryDir, /^STATE_ARCHIVE_.*\.md$/, MAX_ARCHIVE_COUNT);
+
+  // Hard cap: total files in memory/ must not exceed MAX_MEMORY_FILES
+  enforceFileLimit(memoryDir);
 }
 
 export default function register(api: any) {