Memory Recall

You have access to a role-based persistent memory system with 4 layers:

L2 Structured  → memory/consolidated.md (deduplicated, priority-ranked)
PENDING        → memory/pending.md (auto-extracted, awaiting verification)
L1 Raw         → memory/daily/YYYY-MM-DD.md (session logs)
Knowledge      → docs/knowledge/ (reusable patterns, architecture decisions)

Tools Available

Tool	Purpose
role_search({ query: "<text>" })	Search all layers. Auto-reinforces high-score matches (≥0.5). Auto-promotes relevant pending memories.
role_exec({ op: "list" })	List all consolidated memories, detect issues
role_exec({ op: "role_info" })	List the active role directory structure; does not read file contents
role_search({ query: "<text>", scope: "knowledge" })	Search knowledge base

Process

Step 1: Targeted search

role_search({ query: "<user topic or key concept>" })

The search automatically:

• Searches consolidated learnings, preferences, events (block-level milestones)
• Searches pending (all matches ≥ minScore surface as [pending]; score ≥0.5 auto-promotes to learning)
• Searches last 7 days of daily files (EVENT/LESSON/PREFERENCE keep their kind)
• Tag boost: matching tags +0.3 score, related tags +0.15
• Auto-reinforce: matches ≥0.5 get used count +1

Step 2: Scan High Priority

If search returns few results:

role_exec({ op: "list" })

Focus on High Priority [3x]+ — these are battle-tested.

Step 3: Deep context (if needed)

• Narrow the memory.search query, or use knowledge.search for reusable artifacts.
• If the current task explicitly requires a core role file, inspect the exact injected path with the standard file-read tool; do not scan role files as a startup ritual.

Step 4: Check knowledge base

For technical tasks:

role_search({ query: "<topic>", scope: "knowledge" })

Step 5: Summarize and proceed

Summarize findings, then proceed.

Guardrails

• max memory ops: 10 — Don't burn the whole session searching
• Tag boost is real — matching tags rank higher. Trust the sort.
• If nothing found, proceed — not every task has prior knowledge
• Summarize before proceeding

Memory Format

# Learnings (High Priority)    → used ≥ 3
- [6x] 声明完成前验证铁律

# Learnings (Normal)           → used 1-2
- [2x] 软删除优先

# Learnings (New)              → used = 0
- [0x] 标签系统闭环是快速win

# Preferences: Communication | Code | Tools | Workflow | General
- 偏好中文沟通

Pending Layer

Auto-extracted memories land in memory/pending.md:

• [○] pending — awaiting verification
• [✓] promoted — moved to consolidated
• [✗] discarded — 7 days without use

Search auto-promotes pending entries with score ≥0.5. Usage is verification.

Tags

Each learning has LLM-auto-extracted tags. Search uses them:

• Exact tag match → +0.3 score
• Related tag (association graph) → +0.15 score
• This means conceptually related entries surface even with different wording

Important

• Start with targeted role_search({ query: "..." }) only when prior context is materially relevant; search may reinforce or promote matches
• High Priority [3x]+ are most valuable — read first
• User references past work → search for related keywords
• Nothing found → proceed without memory