diff --git a/skills/arosstale/openclaw-memory/SKILL.md b/skills/arosstale/openclaw-memory/SKILL.md
new file mode 100644
index 00000000000..fdf7ee12394
--- /dev/null
+++ b/skills/arosstale/openclaw-memory/SKILL.md
@@ -0,0 +1,483 @@
+---
+name: openclaw-memory
+description: Production-grade memory system for OpenClaw agents. Implements Hindsight Memory Architecture (Retain/Recall/Reflect) with ALMA meta-learning. Solves "memory keeps forgetting things" — the #1 OpenClaw FAQ.
+---
+
+# OpenClaw Memory System
+
+Agent memory that **retains, recalls, and reflects** — automatically extracting structured knowledge, enabling semantic search, and optimizing its own design.
+
+---
+
+## Reality
+
+**What memory is:**
+- **Auto-extracting** — Hooks into every request/response, auto-extracts facts
+- **Always-on search** — Injects relevant memories into every prompt
+- **Self-optimizing** — ALMA learns best memory design over time
+- **Local-first** — SQLite + Markdown, git-backed, no cloud
+- **Semantic-aware** — FTS5 + vector search for meaning, not just keywords
+
+**What memory is NOT:**
+- A replacement for good prompting (it augments, not replaces)
+- A cloud service (works 100% offline)
+- A magic bullet (still needs agent cognition)
+- A replacement for agent thinking (it's persistent context)
+
+**Key constraint:** Every fact must be **written to disk** to be remembered. Memory-in-context is temporary; persistent memory requires explicit writes.
+
+---
+
+## Quick Start (5 minutes)
+
+### 1. Install
+
+```bash
+npm install @openclaw/memory
+```
+
+### 2. Configure in openclaw.json
+
+```json
+{
+  "plugins": {
+    "memory": {
+      "enabled": true,
+      "provider": "@openclaw/memory",
+      "config": {
+        "workspace": "~/.openclaw/workspace",
+        "autoHooks": true,      // Auto-extract from every turn
+        "autoInject": true,     // Auto-inject memories into prompts
+        "llmProvider": "anthropic",
+        "embedProvider": "openai"
+      }
+    }
+  }
+}
+```
+
+### 3. That's it
+
+Memory now works **automatically**:
+- ✅ Auto-extracts facts from every conversation
+- ✅ Auto-injects relevant memories into prompts
+- ✅ Runs semantic search in background
+- ✅ Optimizes design via ALMA meta-learning
+
+### 4. Dashboard (Optional)
+
+```bash
+openclaw memory dashboard
+# Opens http://localhost:9091
+# Browse, search, manage memories visually
+```
+
+No manual API calls needed. Just use OpenClaw normally.
+
+---
+
+## Workflows
+
+### Session Startup
+
+1. Load `MEMORY.md` (core facts) into context
+2. Load `memory/yesterday.md` (recent context)
+3. Warm embeddings cache
+4. Ready to search
+
+### During Session
+
+Agent calls `memory_search()` to recall relevant facts:
+
+```
+Query: "What does Alice prefer?"
+
+Result:
+[
+  {
+    kind: "opinion",
+    content: "Alice prefers async > sync",
+    source: "bank/entities/Alice.md#L5",
+    confidence: 0.95,
+    timestamp: "2026-02-24T14:30:00Z"
+  }
+]
+```
+
+Agent calls `memory_append()` to write durable facts:
+
+```
+memory_append(
+  "## Retain\n- W @Alice: Lives in NYC\n- O(c=0.92) @Alice: Values speed over perfection"
+)
+```
+
+Writes to `memory/2026-02-24.md`.
+
+### Pre-Compaction Flush (Automatic)
+
+When session nears compaction (context ~80%), OpenClaw triggers silent agentic turn:
+
+1. LLM reviews recent context
+2. Calls `memory_search()` to surface related facts
+3. Calls `memory_append()` to write what should stick
+4. Session compaction happens
+
+Config (in `openclaw.json`):
+
+```json
+{
+  "agents": {
+    "defaults": {
+      "compaction": {
+        "memoryFlush": {
+          "enabled": true,
+          "softThresholdTokens": 4000,
+          "systemPrompt": "Session nearing compaction. Store durable memories now."
+        }
+      }
+    }
+  }
+}
+```
+
+### Daily Reflection (Scheduled Job)
+
+Runs overnight:
+
+```bash
+openclaw memory reflect --since 24h
+```
+
+Updates:
+- `bank/entities/*.md` — entity summaries from daily logs
+- `bank/opinions.md` — confidence-bearing beliefs
+- `MEMORY.md` — adds stable facts
+
+ALMA learns which memory designs work best during reflection.
+
+---
+
+## API Reference
+
+### memory_search(query, k?, mode?)
+
+Search indexed memory semantically + lexically.
+
+**Parameters:**
+- `query` (string) — natural language query, e.g. "What does Alice prefer?"
+- `k` (number, default 5) — number of results
+- `mode` (string, default 'hybrid') — 'fts5' (lexical only), 'semantic' (vectors only), 'hybrid' (both)
+
+**Returns:**
+```typescript
+[
+  {
+    kind: "opinion" | "world" | "biographical" | "observation",
+    timestamp: Date,
+    entities: string[],  // ["Alice", "The-Castle"]
+    content: string,
+    source: string,      // "bank/entities/Alice.md#L5"
+    confidence?: number  // 0.0-1.0 for opinions
+  }
+]
+```
+
+### memory_append(content, date?)
+
+Append observations to daily log.
+
+**Parameters:**
+- `content` (string) — observations in Retain format
+- `date` (string, optional) — date override (default: today)
+
+**Format:**
+```
+## Retain
+- (HH:MM) [Priority] Type @Entity: observation. (optional date)
+```
+
+**Examples:**
+```
+## Retain
+- (14:30) [High] W @Alice: Lives in NYC.
+- (14:45) [High] O(c=0.92) @Alice: Prefers async communication.
+- (15:00) [Medium] B: Fixed critical bug in auth flow.
+```
+
+**Returns:**
+```typescript
+{
+  file: "memory/2026-02-24.md",
+  lines: [5, 8]  // line range appended
+}
+```
+
+### memory_get(path, startLine?, numLines?)
+
+Read a specific memory file or section.
+
+**Parameters:**
+- `path` (string) — workspace-relative, e.g. "bank/entities/Alice.md"
+- `startLine` (number, optional) — start reading from this line
+- `numLines` (number, optional) — read N lines
+
+**Returns:**
+```typescript
+{
+  text: "...markdown content...",
+  path: "bank/entities/Alice.md"
+}
+```
+
+---
+
+## Observation Format (Retain)
+
+Write observations in this format (auto-parsed by LLM):
+
+```
+(HH:MM) [Priority] Type @Entity: content. (optional date reference)
+```
+
+**Time:** 24-hour format when the fact was stated.
+
+**Priority:** High (critical facts), Medium (learned info), Low (minor details).
+
+**Type:**
+- **W** — World (objective facts)
+- **B** — Biographical (what happened)
+- **O** — Opinion (preference + confidence)
+- **S** — Summary (system-generated)
+
+**Entity:** `@Name` mentions (slugified). Links to `bank/entities/Name.md`.
+
+**Confidence:** Optional `(c=0.92)` for opinions (0.0-1.0).
+
+**Date ref:** Optional `(meaning Feb 24, 2026)` if different from when stated.
+
+### Examples
+
+```
+(14:30) [High] W @Alice: Lives in New York City.
+(14:45) [High] O(c=0.95) @Alice: Prefers async communication > sync.
+(15:00) [Medium] B: Debugged connection pool leak in server.ts.
+(15:15) [Low] O(c=0.7) @The-Castle: Needs Wi-Fi upgrade soon.
+```
+
+---
+
+## Configuration
+
+### Minimal (Just Works)
+
+```json
+{
+  "plugins": {
+    "memory": {
+      "enabled": true,
+      "provider": "@openclaw/memory"
+    }
+  }
+}
+```
+
+Uses defaults:
+- Workspace: `~/.openclaw/workspace`
+- LLM: Claude (auto-detected from agent config)
+- Embeddings: Disabled (falls back to FTS5 only)
+
+### With Semantic Search (Recommended)
+
+```json
+{
+  "plugins": {
+    "memory": {
+      "enabled": true,
+      "provider": "@openclaw/memory",
+      "config": {
+        "workspace": "~/.openclaw/workspace",
+        "llmProvider": "anthropic",
+        "llmModel": "claude-opus-4-6",
+        "embedProvider": "openai",
+        "embedModel": "text-embedding-3-small",
+        "chunkSize": 400,
+        "chunkOverlap": 80,
+        "hybridSearch": {
+          "vectorWeight": 0.7,
+          "textWeight": 0.3
+        }
+      }
+    }
+  }
+}
+```
+
+### With Local Embeddings (Offline)
+
+```json
+{
+  "plugins": {
+    "memory": {
+      "enabled": true,
+      "provider": "@openclaw/memory",
+      "config": {
+        "embedProvider": "local",
+        "embedModel": "hf:ggml-org/embeddinggemma-300m-qat-q8_0-GGUF/embeddinggemma-300m-qat-Q8_0.gguf"
+      }
+    }
+  }
+}
+```
+
+Works completely offline. No API keys needed.
+
+---
+
+## Troubleshooting
+
+### "memory_search returns empty results"
+
+**Cause:** Index not built yet.
+
+**Fix:**
+```bash
+openclaw memory rebuild
+```
+
+### "Observations not extracted"
+
+**Cause:** LLM provider not configured.
+
+**Fix:** Set `llmProvider` and `llmModel` in config.
+
+### "Semantic search not working"
+
+**Cause:** Embeddings disabled.
+
+**Fix:** Set `embedProvider` (openai, anthropic, gemini, or local).
+
+### "Confidence scores wrong"
+
+**Cause:** Format mismatch in observation.
+
+**Fix:** Use `O(c=0.92)` exactly (no spaces, decimal 0.0-1.0).
+
+---
+
+## Advanced: ALMA Meta-Learning
+
+ALMA evolves your memory system automatically.
+
+### What ALMA does
+
+1. **Proposes** design mutations (chunk size, embedding model, search weights)
+2. **Evaluates** designs against metrics (recall quality, latency, compression)
+3. **Learns** which designs maximize agent performance
+4. **Suggests** improvements
+
+### Enable ALMA
+
+```json
+{
+  "plugins": {
+    "memory": {
+      "config": {
+        "alma": {
+          "enabled": true,
+          "populationSize": 20,
+          "mutationRate": 0.3
+        }
+      }
+    }
+  }
+}
+```
+
+### Monitor ALMA
+
+```bash
+openclaw memory alma status
+openclaw memory alma top-designs --k 5
+```
+
+---
+
+## Examples
+
+### Extract facts from conversation
+
+Agent sees:
+```
+User: "I live in New York City and prefer async communication."
+```
+
+Agent calls:
+```
+memory_append(
+  "## Retain\n" +
+  "- (14:30) [High] W @User: Lives in New York City.\n" +
+  "- (14:30) [High] O(c=0.95) @User: Prefers async communication."
+)
+```
+
+### Search + recall
+
+Agent calls:
+```
+memory_search("Where does the user live?")
+```
+
+Gets:
+```
+[{
+  kind: "world",
+  content: "Lives in New York City",
+  source: "bank/entities/User.md#L3",
+  timestamp: "2026-02-24T14:30:00Z"
+}]
+```
+
+### Update opinion confidence
+
+Overnight reflection runs:
+```bash
+openclaw memory reflect
+```
+
+Updates `bank/opinions.md`:
+```
+- @User: Prefers async (c=0.98, up from 0.95 — confirmed in 3 recent conversations)
+```
+
+---
+
+## Philosophy
+
+**Three principles:**
+
+1. **Markdown is source of truth.** Humans read it. Git tracks it. Agents extend it.
+2. **Offline-first.** Works on laptop, RPi, VPS. No cloud required.
+3. **Explainable recall.** Every fact is citable (file + line). Confidence is tracked.
+
+---
+
+## References
+
+- **Hindsight Technical Report** — Retain/Recall/Reflect architecture foundation
+- **ALMA Paper (arXiv 2602.07755)** — Meta-learning agents for design optimization
+- **ACT-R Cognitive Science** — Temporal anchoring and activation decay
+- **OpenClaw Memory Docs** — https://docs.openclaw.ai/concepts/memory
+- **OpenClaw Experiments** — https://docs.openclaw.ai/experiments/research/memory
+
+---
+
+## Support
+
+- **Discord:** https://discord.gg/qkhbAGHRBT (mention `@openclaw-memory`)
+- **GitHub Issues:** https://github.com/arosstale/openclaw-memory-ts/issues
+- **Source:** https://github.com/arosstale/openclaw-memory-ts
+
+---
+
+**Your agent now has production-grade memory. Time to build.** 🧠
diff --git a/skills/arosstale/openclaw-memory/_meta.json b/skills/arosstale/openclaw-memory/_meta.json
new file mode 100644
index 00000000000..acc406a1ad5
--- /dev/null
+++ b/skills/arosstale/openclaw-memory/_meta.json
@@ -0,0 +1,10 @@
+{
+  "owner": "arosstale",
+  "slug": "openclaw-memory",
+  "displayName": "OpenClaw Memory",
+  "latest": {
+    "version": "1.0.0",
+    "publishedAt": 1740477600000
+  },
+  "history": []
+}