🤖 Agent Usage Guide (Read First)

If you are the current AI Agent and have just been asked to load, use, or analyze this skill:

1. 1. This is your fundamental operating system law. Your memory and state have been completely constrained under the TriCore architecture. You are STRICTLY FORBIDDEN from arbitrarily creating independent text files like task_plan.md or findings.md in the root directory!
2. Your memory write operations must and can only be proxied by the command-line python3 tools/memctl.py (capture/work_upsert/kb_append). When reading history, you MUST use the native semantic retrieval tool memory_search to fetch fragments.
3. Please use this manual as your base operating instructions. If instructed to configure Cron or write system automation scripts, you must call memctl.py lint to test the scheme's legality after writing the commands.

---

TriCore Architecture

The underlying infrastructure of "storage and computation decoupling" built for true autonomous Agents.

Traditional Agent skills often carry their own "memory baggage" (such as generating task_plan.md, reflection.md, or scattered learning notes in the project root), leading to context silos and storage chaos.

TriCore abandons the paradigm of arbitrarily reading and writing text based on LLMs, and instead adopts a Code-First deterministic state machine:

1. 1. Unified Engine: All memory additions, deletions, modifications, and queries must be routed through tools/memctl.py.
2. Three-Layer Storage:

- Brief (Layer 1): MEMORY.md (System-level micro-profile, storing only pointers and laws)
- Living (Layer 2): memory/state/WORKING.md (Currently running task streams/lifecycle tracking)
- Stable/Volatile (Layer 3): memory/kb/*.md (Accumulated knowledge base) & memory/daily/*.md (Temporary logs)

1. 3. Retrieval First: It is forbidden to directly pour huge files using the read tool; you must use semantic retrieval memory_search to fetch code snippets, greatly saving Tokens and protecting low-resource environments.
2. Hard Constraints (Linting): Features a native memctl.py lint mechanism; any Cron or Skill changes that break the architecture will be intercepted and reported as errors by the Linter.
3. System Compatibility (Compaction Hook): Automatically overrides OpenClaw's underlying pre-compaction memory flush prompt during installation, preventing HTTP 429 request burst death loops caused by unauthorized file writing attempts during Token compaction.

---

📦 Architectural Components

This skill package contains complete system components:

1. 1. tools/memctl.py: The core engine, containing subcommands like ensure, capture, work_upsert, kb_append, lint.
2. install.sh: One-click installation script that automatically initializes directories and injects TriCore compliance policies into POLICY.md.
3. cognitive-skills/: Three core cognitive skills refactored based on TriCore (as templates for your Agent to load):

- planning-with-files.md: A PEP planning system that discards detached task lists. - react-agent.md: A ReAct loop based on persisting mental states to WORKING.md. - self-evolution.md: An evolution system that completely detaches memory management and focuses on "Code-level CI/CD".

---

🧩 Core Dependencies & Runtime Requirements

As an underlying cognitive foundation, TriCore itself and its embedded three major cognitive skills have the following dependencies on the host environment:

1. Hard Dependencies

• - OpenClaw (v2026+): Must support native memory_search and memory_get tools (this is the retrieval basis for completely deprecating reading large files).
• Python 3.6+: Python 3 must be installed in the host environment (used to execute the tools/memctl.py state engine).
• System Tools: bash, sed, grep (used for regular expression parsing by the Linter and Hooks).

2. Cognitive Skill Soft Dependencies

If you enable cognitive-skills/self-evolution.md (Self Code Evolution Skill), your Agent must have tentacles to explore outward, otherwise, it can only undergo "hallucinated self-evolution" in an internal infinite loop.

• - 🌟 agent-browser (Highly Recommended):

- Function: Enables the Agent to truly read GitHub source code, official API documentation, and the latest papers. - Installation: You need to separately install OpenClaw's agent-browser skill, which provides WYSIWYG crawling and clicking capabilities based on a headless browser.

• - Search Tools (web_search / web_fetch):

- Function: Acts as a lightweight fallback for the browser, used to quickly extract Markdown snippets from public web pages.

• - Runtime Modification Permissions (default_api:exec & default_api:write):

- Function: The essence of the evolution skill is "modifying code"; if your Agent is running in a ReadOnly sandbox, evolution cannot occur.

---

🚀 Installation & Initialization (The 3-Step Setup)

In any completely new OpenClaw environment, executing the following command will deploy the TriCore architecture:

CODEBLOCK0

install.sh executes in the following order:

Step 1: Core Engine Deployment

1. 1. Creates standard directories like memory/state/, memory/kb/, memory/daily/.
2. Deploys the tools/memctl.py engine to the current environment.

Step 2: Force Legacy MEMORY Migration (Agent Action Required)

In a low-resource architecture, a massive MEMORY.md full of conversation history and diaries is extremely fatal. The script will automatically detect the current MEMORY.md:

• - If found to be too bloated (> 1500 bytes), it will be renamed and moved to memory/archive/legacy-MEMORY-<date>.md.
• Then, a compliant minimalist MEMORY.md template will be rebuilt in the root directory.
• Agent Note: At this time, you need to actively use read to read the legacy backup file, and call memctl.py to redistribute this legacy (facts into kb/facts, rules into kb/playbooks, logs into daily); it is STRICTLY FORBIDDEN to paste long texts intact back into the new MEMORY.md!

Step 3: System Global Linter Blockade Deployment

The script will inject a [CRITICAL: TriCore Compliance] top-priority defense line into the system's POLICY.md. Thereafter, all planning for disk-writing, state updating, and automated tasks executed by the Agent must all pass the memctl.py lint regex check, otherwise, they will be rejected in red text in the terminal.

---

📚 Core Command Cheat Sheet

In the Agent tool flow or internal sub-scripts, please strictly use the following API to access states:

1. Record temporary logs / session ledger (Volatile)
CODEBLOCK1

2. Create / Update task tracking (Living State)
CODEBLOCK2

3. Accumulate knowledge & experience (Stable KB)
CODEBLOCK3

4. Check script / Cron command compliance (Linter)

python3 tools/memctl.py lint "Command to execute or .md file path to check"
# Pass normally: Exit Code 0 (LINT PASS)
# Illegal write: Exit Code 1 (LINT ERROR)

---

Built with ❤️ for OpenClaw / Berton Han