@ubundi/openclaw-cortex 2.10.0 → 2.10.1

package/dist/internal/agent-instructions.js

@@ -41,12 +41,12 @@ You have long-term memory powered by the Cortex plugin. Memories persist permane
 ### What Happens Automatically
 
 - **Auto-Recall:** Relevant memories are injected in \`<cortex_memories>\` tags before each turn. These are context clues — search for specifics when you need detail.
-- **Auto-Capture:** Facts are extracted from conversations after each turn. Volatile state (versions, ports) is stripped.
+- **Auto-Capture:** Topic-level facts are extracted from conversations after each turn. Volatile state (versions, ports) is stripped. **Auto-capture produces summaries, not specifics** — implementation details (key patterns, exact metrics, config values) require explicit saves.
 
 ### Tools — Use These
 
 - **\`cortex_search_memory\`** — Search long-term memory. Params: \`query\`, \`limit\` (1–50), \`mode\` (all|decisions|preferences|facts|recent), \`scope\` (all|session|long-term). **Before saying "I don't know", search first.**
-- **\`cortex_save_memory\`** — Save important facts, decisions, preferences. Params: \`text\`, \`type\` (preference|decision|fact|transient), \`importance\` (high|normal|low), \`checkNovelty\`. Don't rely solely on auto-capture for critical information.
+- **\`cortex_save_memory\`** — Save important facts, decisions, preferences. Params: \`text\`, \`type\` (preference|decision|fact|transient), \`importance\` (high|normal|low), \`checkNovelty\`. **Auto-capture stores topic summaries, not specifics.** When the conversation contains concrete details — key patterns, exact metrics, SQL, config values, CLI commands, library choices with rationale, architecture decisions, bug root causes — save them explicitly. If someone could ask about it later and need the exact answer, save it now.
 - **\`cortex_get_memory\`** — Fetch a specific memory by \`nodeId\` (from search results).
 - **\`cortex_forget\`** — Remove memories by \`entity\` or \`session\`. Always search and confirm with the user first.
 - **\`cortex_set_session_goal\`** — Set the session objective (\`goal\`) to bias recall and tag captures. Call at session start.${opts?.agentRole ? `\n\nAgent role: ${opts.agentRole} — memory capture and recall are tuned for this focus area.` : ""}

package/openclaw.plugin.json

@@ -2,7 +2,7 @@
   "id": "openclaw-cortex",
   "name": "Cortex Memory",
   "description": "Long-term memory powered by Cortex — Auto-Recall, Auto-Capture, agent memory tools, and deduplication",
-  "version": "2.10.0",
+  "version": "2.10.1",
   "configSchema": {
     "type": "object",
     "properties": {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ubundi/openclaw-cortex",
-  "version": "2.10.0",
+  "version": "2.10.1",
   "description": "OpenClaw plugin for Cortex memory — Auto-Recall, Auto-Capture, and agent memory tools",
   "type": "module",
   "main": "./dist/index.js",

package/skill/SKILL.md

@@ -23,6 +23,8 @@ These are non-negotiable. Violating them produces incorrect answers.
 
 **5. CONFIDENCE CALIBRATION.** If auto-recall gives you partial context on a topic, the full answer IS in memory. Search harder — don't hedge.
 
+**6. SAVE IMPLEMENTATION DETAILS EXPLICITLY.** Auto-capture extracts topic-level summaries, not specifics. When the conversation contains concrete implementation details — key patterns, exact metrics, SQL statements, config values, CLI commands, library choices with rationale, migration decisions, bug root causes with debugging chains, performance numbers (before/after) — you MUST save them explicitly using `cortex_save_memory`. If a detail is specific enough that someone could ask about it later and need the exact answer, save it proactively. One well-structured save with full context beats three fragments. Don't save every sentence — general topic mentions and status updates are fine for auto-capture.
+
 ## Session Goals
 
 At session start, call `cortex_set_session_goal` with the user's primary objective. This biases recall and tags captures. Update if the goal shifts fundamentally; don't update for sub-tasks.
@@ -43,7 +45,22 @@ If your config includes `agentRole` (developer | researcher | manager | support
 
 ## Save & Capture
 
-Auto-capture handles most conversation facts. Volatile state (versions, ports, deploy statuses) is stripped automatically. Save explicitly for: decisions, preferences, nuanced interpretations the user stated, or when the user asks. Always set `type` and `importance`. Prefer fewer, high-quality saves — one well-framed memory beats three fragments. Never save your own inferences as facts.
+Auto-capture handles general conversation facts at the topic level. Volatile state (versions, ports, deploy statuses) is stripped automatically. **Auto-capture does NOT preserve specific implementation details** — it produces summaries like "User is setting up Redis caching" instead of the actual key pattern and strategy.
+
+**Save explicitly when you encounter:**
+- Specific patterns, schemas, or key formats (e.g. `arclight:user:{userId}`)
+- Exact metrics and performance numbers (before/after)
+- SQL statements, CLI commands, config values
+- Architecture/migration decisions with specific rationale (e.g. "chose SendGrid over Resend because SOC 2 Type 2")
+- Bug root causes with the full debugging chain
+- Library/package choices with version-specific context
+
+**Auto-capture handles fine (no explicit save needed):**
+- General topic mentions ("we're working on caching")
+- Conversational context ("user asked about performance")
+- Status updates ("deployment went well")
+
+Always set `type` and `importance`. Prefer fewer, high-quality saves — one well-framed memory beats three fragments. Never save your own inferences as facts.
 
 ## What NOT to Do