Nox-Lumen Mfgself-awareness
Capabilities
self-awareness is the Agent’s self-model handbook. After context compaction, the Agent’s “memory” may be lossy — this skill lets the Agent re-read authoritative architecture facts and behave correctly.
Post-compaction “memory” is not ground truth. Use unified_search(action="get_round") or search to fetch exact payloads before asserting facts — never treat compaction summaries as verbatim truth.
Problems it solves
| Scenario | Without self-awareness | With self-awareness |
|---|---|---|
| After compaction, user asks “what was X?” | Guess from fuzzy memory | Retrieve original output first |
| Session resumes mid-run | Unsure where to continue | Recover from Plan appendix |
| Unsure whether an action is allowed | Trial-and-error or guessing | Consult system boundaries first |
| User asks what the Agent can do | Rough guesswork | Read installed skills explicitly |
Three subsystems to introspect
self-awareness clarifies three pillars:
| Subsystem | What it does | What you can extend |
|---|---|---|
| Memory | unified_search for history, preferences, KB | System accumulates discoveries into LTM |
| Tools | execute_code / unified_search / write_file / bash, etc. | Register reusable scripts via ScriptRegistry |
| Skills | Installed skills (see <skills> list) | Extend via skills_search / skills_install |
Core principle: “search before act”
The Agent has unified_search to query execution history, user docs, prior artifacts, and cross-session memory. Whenever uncertain (prior results, user intent, file contents), search first — do not guess.
How to trigger
This skill is marked always-inject — it is injected automatically on every session start. Users need not invoke it.
You may still invoke explicitly:
Related concepts
- Core concepts · Memory — compaction mechanics
- core · Context compaction
- memory-sdk