智能体 | Nanobot 上下文管理介绍

def build_system_prompt(self, skill_names=None, channel=None):
    parts = []
    
    parts.append( self._get_identity(channel=channel) )           # Part 1: 身份标识
    parts.append( self._load_bootstrap_files() )                   # Part 2: 引导文件
    parts.append( f"# Memory\n\n{memory}" )                        # Part 3: 长期记忆
    parts.append( f"# Active Skills\n\n{always_content}" )         # Part 4: 常驻技能
    parts.append( render_template("skills_section.md", ...) )   # Part 5: 技能目录
    
    return "\n\n---\n\n".join(parts)  # 用 --- 分隔各段

MemoryStore.get_memory_context()
  → read_memory()
    → read_file(workspace/memory/MEMORY.md)   # 返回纯文本
  → 返回 "## Long-term Memory\n{content}"

def build_messages(self, history, current_message, media, channel, chat_id, ...):
    # 第一部分：运行时上下文
    runtime_ctx = self._build_runtime_context(channel, chat_id, self.timezone)
    #  → "[Runtime Context — metadata only, not instructions]
    #     Current Time: 2026-05-22 15:30 CST
    #     Channel: telegram
    #     Chat ID: 12345"

    # 第二部分：用户实际输入（文本 or 文本+图片）
    user_content = self._build_user_content(current_message, media)
    #  → "帮我写一段 Python 代码" 
    #    或 [{"type":"image_url",...}, {"type":"text","text":"这张图是什么"}]

    # 合并为单条 user message
    merged = f"{runtime_ctx}\n\n{user_content}"
    # 或多模态: [{"type":"text", "text": runtime_ctx}, image_block, text_block]

@staticmethod
def _build_runtime_context(channel, chat_id, timezone):
    lines = [f"Current Time: {current_time_str(timezone)}"] 
    if channel and chat_id:
        lines += [f"Channel: {channel}", f"Chat ID: {chat_id}"]
    return "[Runtime Context — metadata only, not instructions]\n" + "\n".join(lines)

messages = [
    {"role": "system", "content": "完整 system prompt..."},          ← 系统指令
    {"role": "user",   "content": "..."},                      ← 历史消息 N-2
    {"role": "assistant", "content": "..."},                   ← 历史消息 N-1
    {"role": "user",   "content": runtime_ctx + 用户输入},      ← 当前提问 (合并)
]

用户对话消息
    │
    ▼ (实时)
session.messages[] ──→ [Token 超限?] ──→ Consolidator.archive()
                                            │
                                            ▼
                                    history.jsonl (append-only 日志)
                                            │
    ┌───────────────────────────────────────┘
    │
    ▼ (定期/demand)
Dream.run()
    │
    ├─→ Phase 1: LLM 分析历史 + 对比当前记忆文件
    │
    ├─→ Phase 2: LLM 用 edit_file 工具修改
    │         ├── SOUL.md      (更新人格)
    │         ├── USER.md      (更新画像)
    │         └── MEMORY.md    (更新知识)
    │
    └─→ GitStore.auto_commit() (版本快照)

下次构建 System Prompt 时:
    context.py 读取最新的 SOUL/USER/MEMORY
    → 注入 Part 2 (Bootstrap) + Part 3 (Memory)
    → Agent 拥有了跨会话的持久记忆 

Compare conversation history against current memory files. 
Also scan memory files for stale content.

Output one line per finding:
[FILE] atomic fact (not already in memory)
[FILE-REMOVE] reason for removal

Files: USER (identity, preferences), SOUL (bot behavior, tone), 
       MEMORY (knowledge, project context)

Rules:
- Atomic facts: "has a cat named Luna" not "discussed pet care"
- Corrections: [USER] location is Tokyo, not Osaka

Staleness — flag for [FILE-REMOVE]:
- Time-sensitive data older than 14 days: weather, daily status...
- Completed one-time tasks: finished research, resolved incidents
- Superseded: approaches replaced by newer solutions

Do not add: current weather, transient status, temporary errors.

[SKIP] if nothing needs updating.

[MEMORY] User prefers Chinese communication (not English)
[MEMORY] Project uses Deepseek as primary model
[MEMORY] Debug launch config created at .vscode/launch.json for onboard/agent/telegram
[USER] Language: 中文 (Chinese)
[USER] Tech stack: Python 3.12+, macOS arm64, VSCode
[USER] Editor: VSCode with Python debugging
[SOUL-REMOVE] "Use formal tone in all responses" — user prefers concise casual style based on interaction pattern
[SOUL] Be concise but thorough. Use short paragraphs for messaging apps.
[SKIP] weather/status transient data ignored

Update memory files based on the analysis below.
- [FILE] entries: add content to appropriate file
- [FILE-REMOVE] entries: delete from memory files

File paths: SOUL.md, USER.md, memory/MEMORY.md

Editing rules:
- Edit directly — file contents provided below
- Use exact text as old_text, include surrounding blank lines
- Batch changes to same file into one edit_file call
- Surgical edits only — never rewrite entire files
- If nothing to update, stop without calling tools

tracked_files = ["SOUL.md", "USER.md", "memory/MEMORY.md"]

# .gitignore 策略：
# /*                    ← 忽略所有文件
# !memory/              ← 但跟踪 memory 目录
# !SOUL.md              ← 跟踪 SOUL.md
# !USER.md              ← 跟踪 USER.md
# !.gitignore

sha = self.store.git.auto_commit(f"dream: 2026-05-09 15:30, 4 change(s)")
# commit author: "nanobot <nanobot@dream>"