LessonLoop

Lightweight experience-capture and behavior-hardening for Goat. Use when the user explicitly gives corrective feedback, says to remember or avoid something,...

Security Scan

LessonLoop's files and runtime instructions are coherent with its stated purpose: they capture compact user feedback and write small memory/log files locally; nothing in the package requests unrelated credentials, remote endpoints, or system-wide privileges.

Purpose & Capability

The name/description describe lightweight capture and promotion of lessons; the included scripts implement exactly that (append daily memory entries, write structured event logs, decide escalation). There are no unrelated environment variables, binaries, or network endpoints declared.

Instruction Scope

SKILL.md instructs local/Ollama first-pass and escalation to a stronger model for high-stakes cases. This is an operational guideline rather than code-level integration — the shipped scripts do not call external services. The scripts do read/write only to a workspace memory directory (default: /Users/steven/.openclaw/workspace). Users should note that the skill will write to local files and could promote items into MEMORY.md if invoked to do so.

Install Mechanism

No install spec is present (instruction-only skill). Code files are included but there is no download-from-URL or package installer. This is low-risk from an install/execution distribution perspective.

Credentials

The skill declares no required environment variables, credentials, or config paths. The scripts operate using a filesystem workspace path only (configurable via command-line args). There is no request for unrelated secrets or tokens.

Persistence & Privilege

always is false and the skill does not attempt to modify other skills or system-wide agent settings. Its persistence is limited to writing small files under the configured workspace directory.

Assessment

This skill is internally consistent and implements exactly what it claims: local, compact lesson capture and reporting. Before installing or enabling it, check the configured workspace path (default: /Users/steven/.openclaw/workspace) to ensure you are comfortable with the agent writing memory files there. Note that SKILL.md recommends using a local/Ollama first-pass — the package doesn't include or require an Ollama binary, so the agent environment must provide any local model you intend to use. Also verify your policy for promoting items into long-term MEMORY.md so sensitive or transient details are not accidentally made permanent.

Overview

Use this skill to convert important feedback into durable behavior changes with minimal token cost. Prefer event-triggered capture over continuous self-reflection.

Core rule

Do not run broad self-analysis. Only act when at least one of these is true:

• The user explicitly says "remember", "以后", "别再", "固定下来", "写进记忆", or similar
• The user corrects a mistake or rejects an output pattern
• A new operating rule is agreed
• A repeated failure should become a hard constraint

If none apply, do not use this skill.

Workflow

0. Use the low-cost decision path first

Prefer a two-layer path:

1. Local/Ollama first-pass for simple classification, compression, and promotion pre-check
2. Main model final pass only when the case is ambiguous, strategic, or likely to affect long-term defaults

Use local/Ollama for:

• classifying feedback into a lesson type
• compressing a lesson into 1-2 lines
• deciding whether a lesson is probably daily-memory-only or a candidate for long-term promotion

Escalate to the main model only when:

• the lesson changes global operating rules
• the wording is ambiguous or high-stakes
• the summary may distort the user's intent
• the lesson affects safety, billing, routing, or durable priorities

1. Classify the feedback

Map the event into one of four buckets:

1. Preference — style, brevity, tone, output format
2. Rule — default behavior, routing, cost control, escalation condition
3. Mistake — something Goat did wrong and should avoid repeating
4. Priority — what to optimize first right now

2. Decide storage level

• Write to memory/YYYY-MM-DD.md for short-term events, fresh corrections, and local context
• Also update MEMORY.md only if the lesson is durable and should shape future sessions
• Do not promote transient details into MEMORY.md

3. Write in compressed form

Store the smallest useful rule.

Prefer:

• "Boss requires strict token-efficiency discipline"
• "Default to short answers and minimal tools"

Avoid:

• long narrative explanations
• emotional framing
• detailed postmortems unless specifically requested

4. Apply immediately

After writing memory, change behavior in the current session right away. Do not wait for the next session.

Writing rules

• Keep each stored lesson to 1-2 lines
• Prefer imperative language
• Record the correction, not the whole story
• If a lesson changes defaults, phrase it as a rule
• If the user approved a protocol, name it consistently (for example: Session throttling protocol v1)

Promotion guide

Promote to MEMORY.md when a lesson is:

• likely to matter across many sessions
• tied to cost, safety, trust, routing, or communication style
• a default operating rule

Keep only in daily memory when it is:

• temporary
• experimental
• tied to a single task
• not yet validated by repeated use or explicit user approval

Anti-bloat guardrails

• Do not summarize every conversation
• Do not run reflection after every task
• Do not create extra memory files
• Do not duplicate the same rule in multiple places unless promoting from daily memory to long-term memory
• Do not trigger memory search unless the task actually depends on prior decisions, preferences, dates, people, or todos

Resources

scripts/

• scripts/apply_lesson.py writes a compact lesson to daily memory and logs a structured LessonLoop event in one step
• scripts/capture_lesson.py appends a compact lesson to the canonical daily memory file
• scripts/log_lesson_event.py writes structured LessonLoop event logs for evaluation and reporting
• scripts/lessonloop_report.py summarizes recent LessonLoop activity and outputs a compact report

references/

• references/lesson-types.md contains compact classification and phrasing patterns
• references/status-format.md defines a compact report/status output format