菜单
Run a periodic memory evolution cycle: decay stale observations, deprecate low-confidence memories, and prune excess per object. Call every 10 turns or on a daily schedule to keep the memory store healthy and relevant.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Run a multi-agent debate around contested memories. Multiple agents argue for/against a memory claim, then vote to determine its final social status. This is the "game theory enhances collective memory" engine — consensus emerges through structured argumentation, not just majority counting.
Before generating an AI reply, recall relevant memories from DeepMemory and search related code from FlashMemory, then inject both as augmented context. This is the memory-enhanced RAG pipeline — combining symbolic recall with vector search for richer, history-aware responses.
Extract symbolic order from conversations: identify key concepts as Signs, anchor them to ObjectRefs, and record interpretant claims via DeepMemory MCP. This is the "semiotics deconstructs semantics" engine — it turns natural language into structured, contestable memory units instead of opaque vectors.
Process user feedback on AI memories: agreement promotes status, disagreement deletes, refinement revises claims. Translates natural user reactions into DeepMemory governance operations.
FlashMemory CLI for cross-language code analysis and semantic search. Use this skill when you need to understand project architecture, locate implementations, or index large codebases using natural language queries. Make sure to use this skill whenever the user asks about the codebase structure, finding code, or searching files.
Build and maintain semantic vector indexes for large codebases using FlashMemory. Use this skill whenever you need to index a new project, re-index after making code changes, perform incremental file updates, or switch between vector engine backends. Make sure to use this whenever you modify code or pull new changes so the semantic search index stays up-to-date.
| name | memory-evolution-cycle |
| description | Run a periodic memory evolution cycle: decay stale observations, deprecate low-confidence memories, and prune excess per object. Call every 10 turns or on a daily schedule to keep the memory store healthy and relevant. |
| metadata | {"version":"1.0.0","category":"cognitive-memory","tags":["evolution","decay","memory-management","deepmemory"]} |
This skill triggers DeepMemory's built-in evolution engine to automatically maintain memory health. Without periodic evolution, memories accumulate indefinitely — stale observations pollute recall, and low-confidence hints waste context window budget.
DeepMemory MCP Server must be running and have an active memory store.
Call the evolution tool with appropriate parameters:
deepmemory_evolve(
min_confidence=0.35,
stale_after_hours=168,
decay_confidence_delta=0.05,
max_memories_per_object=20
)
Parameter tuning guide:
| Parameter | Conservative | Default | Aggressive |
|---|---|---|---|
min_confidence | 0.20 | 0.35 | 0.50 |
stale_after_hours | 336 (2 weeks) | 168 (1 week) | 72 (3 days) |
decay_confidence_delta | 0.02 | 0.05 | 0.10 |
max_memories_per_object | 50 | 20 | 10 |
The evolution tool returns a structured report:
{
"scanned_count": 142,
"revised_count": 8,
"deleted_count": 3,
"actions": [
{
"memory_id": "mu_abc",
"object_id": "user:preference:editor",
"action": "decay",
"reason": "stale (192h without verification)",
"before_confidence": 0.45,
"after_confidence": 0.40
}
]
}
Flag if any of these conditions are true:
For deeper inspection, check the event log:
deepmemory_events(limit=20)
Look for memory.deleted and memory.revised events to verify the
evolution acted correctly.
| Context | Frequency | Parameters |
|---|---|---|
| Active coding session | Every 10 turns | Conservative |
| Daily background job | Once per day | Default |
| Pre-deployment review | Manual trigger | Aggressive |
| Low-activity project | Weekly | Conservative |
Evolution cycle complete:
Scanned: 142 memories
Decayed: 8 (confidence reduced by 0.05 each)
Pruned: 3 (exceeded max per object)
Deleted: 2 (confidence below 0.35)
Healthy: 129 (no action needed)