← Back to Skills Marketplace
393
Downloads
0
Stars
3
Active Installs
7
Versions
Install in OpenClaw
/install metacognition
Description
Self-reflection engine for AI agents. Extracts patterns from session transcripts into a weighted graph with Hebbian learning and time decay. Compiles a token...
Usage Guidance
This skill appears to do exactly what it says: it will read conversation transcripts (when you instruct the agent to do so), analyze them with the bundled Python script, and persist insights to workspace/memory/metacognition.json and scripts/metacognition-lens.md. Before installing, decide whether you are comfortable with transcripts and derived insights being stored in your workspace. If you enable embeddings, ensure EMBEDDINGS_URL points to a trusted local service (the code only allows localhost addresses). Note the small default-URL mismatch between README/SKILL.md and the script — verify EMBEDDINGS_URL if you rely on embeddings. If you do not want persisted insights, do not enable cron integration or running the script; otherwise the behavior is proportionate to the skill's purpose.
Capability Analysis
Type: OpenClaw Skill
Name: metacognition
Version: 1.1.2
The metacognition skill is a legitimate self-reflection engine that implements Hebbian learning and graph-based insight clustering for AI agents. The code in metacognition.py includes proactive security measures, such as strictly validating that the optional embeddings endpoint is restricted to localhost (127.0.0.1/::1) to prevent data exfiltration and enforcing a 1MB limit on file reads. It relies solely on the Python standard library, avoids dangerous execution functions like subprocess or eval, and the instructions in SKILL.md are entirely aligned with the stated purpose of analyzing session transcripts for behavioral patterns.
Capability Assessment
Purpose & Capability
Name/description (self-reflection, insight graph, lens) align with what the skill installs and requests: a Python script, local storage under workspace/memory, and an optional local embeddings endpoint. The only required binary is python3, which is appropriate.
Instruction Scope
SKILL.md tells the agent to read session transcripts (sessions_list + sessions_history) and to run the included Python CLI to add/reweave/compile insights. Reading conversation history and writing to memory/metacognition.json is expected for this purpose — users should be aware that session text will be analyzed and persisted. The instructions are otherwise scoped to the stated task (no commands that read unrelated system config or post data externally).
Install Mechanism
No install spec (instruction-only + bundled script) — nothing is downloaded from external hosts. The provided Python file is stdlib-only and operates on local files and an optional local HTTP embeddings endpoint.
Credentials
The skill requests no credentials and only uses two optional environment variables (WORKSPACE and EMBEDDINGS_URL), both justified by the design. The script enforces that EMBEDDINGS_URL must resolve to localhost/127.0.0.1/::1 (otherwise embeddings are disabled), which prevents remote exfiltration via embeddings. Minor inconsistency: SKILL.md metadata mentions a default embeddings URL of http://localhost:4821, while metacognition.py defaults to http://localhost:11434 — this is a small mismatch but not a security issue.
Persistence & Privilege
always is false and the skill is user-invocable (normal). It writes to its own workspace files (memory/metacognition.json and scripts/metacognition-lens.md), which match the declared writablePaths in SKILL.md metadata. It does not request system-wide or other-skill configuration changes.
How to Use
- Make sure OpenClaw is installed (local or Docker)
- Run the install command in chat:
/install metacognition - After installation, invoke the skill by name or use
/metacognition - Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.1.2
Enforce localhost-only EMBEDDINGS_URL validation at import time — remote URLs now disable embeddings entirely
v1.1.1
Replace subprocess/curl with stdlib urllib.request — code now matches security metadata claims
v1.1.0
Republish with corrected listing
v1.0.3
Rescan: all curl/subprocess removed in v1.0.2, stdlib urllib only, localhost-only validation
v1.0.2
Security: replaced curl/subprocess with Python stdlib urllib (eliminates external binary dependency and attack surface), added homepage and author, confirmed localhost-only EMBEDDINGS_URL validation
v1.0.1
Security fixes: localhost-only EMBEDDINGS_URL validation (rejects remote endpoints), declared writable/readable paths and env vars in metadata
v1.0.0
Initial release: self-reflection engine with Hebbian learning and graph connections
Metadata
Frequently Asked Questions
What is Metacognition?
Self-reflection engine for AI agents. Extracts patterns from session transcripts into a weighted graph with Hebbian learning and time decay. Compiles a token... It is an AI Agent Skill for Claude Code / OpenClaw, with 393 downloads so far.
How do I install Metacognition?
Run "/install metacognition" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.
Is Metacognition free?
Yes, Metacognition is completely free, licensed under MIT-0. You can download, install and use it at no cost.
Which platforms does Metacognition support?
Metacognition is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).
Who created Metacognition?
It is built and maintained by Mei Park (@meimakes); the current version is v1.1.2.
More Skills