feat(memory): add Hindsight conversation memory

[hotwa]

中文说明

概要

本 PR 为 CyberVerse 增加 Hindsight 长期记忆接入，覆盖 PersonaAgent 数字人对话路径和 Go standard conversation path。

主要变化：

• 增加 Hindsight retain/recall 客户端配置
• 支持动态 bank id：cv:user:{user_id}:character:{character_id}
• 为记忆写入增加 tags、metadata、document_id
• PersonaAgent 在回复前自动 recall，并通过 response_instructions 注入长期记忆上下文
• PersonaAgent 在最终回复后异步 retain 用户/助手对话
• Go standard pipeline 也接入同一套 Hindsight conversation memory
• 更新 README、.env.example、GPU 部署配置和开发计划文档
• 增加 Python/Go 单元测试覆盖 Hindsight payload 和 PersonaAgent memory flow

当前验证状态

CyberVerse 侧接入和写入链路已验证通过：

• PersonaAgent 能正常生成回复
• retain 能写入 Hindsight bank/document
• 写入的 document 包含预期的 transcript、tags、metadata、document_id
• 本地单元测试和 Go orchestrator 测试通过

但端到端“长期记忆可 recall”尚未完全成功。

远程 Hindsight 服务 http://w4ce.jmsu.top:8888 当前 retain/reprocess operation 能完成，但生成结果为：

• memory_unit_count=0
• total_nodes=0
• usage.total_tokens=0
• recall 返回空结果

这说明 CyberVerse 的写入 payload 已进入 Hindsight，但 Hindsight 后端抽取/LLM worker 目前没有生成 memory units。该问题需要在 Hindsight 服务侧继续排查。

测试

已执行：

git diff --check
pixi run -e macos-arm pytest tests/unit/test_hindsight_memory.py tests/unit/test_persona_agent_plugin.py -q
pixi run -e macos-arm go test ./internal/orchestrator
pixi run -e macos-arm python - <<'PY'
import yaml
from pathlib import Path
for path in ["infra/cyberverse_config.example.yaml", "deploy/gpu/cyberverse_config.gpu.yaml"]:
    yaml.safe_load(Path(path).read_text())
PY

结果：

• Python focused tests: 26 passed
• Go orchestrator tests: passed
• YAML parse: passed
• git diff --check: passed

---

English

Summary

This PR adds Hindsight long-term conversation memory support to CyberVerse, covering both the PersonaAgent digital-human path and the Go standard conversation path.

Main changes:

• Add Hindsight retain/recall client configuration
• Support dynamic bank IDs: cv:user:{user_id}:character:{character_id}
• Add tags, metadata, and document IDs to retained memory items
• Automatically recall Hindsight memory before PersonaAgent responses and inject it through response_instructions
• Retain user/assistant turns asynchronously after final PersonaAgent responses
• Wire the Go standard pipeline to the same Hindsight conversation memory client
• Update README, .env.example, GPU deployment config, and the development plan
• Add Python/Go tests for Hindsight payloads and PersonaAgent memory flow

Current Verification Status

The CyberVerse integration and write path have been verified:

• PersonaAgent can respond normally
• retain writes documents into the expected Hindsight bank
• stored documents contain the expected transcript, tags, metadata, and document ID
• local Python and Go tests pass

However, full end-to-end durable recall is not yet successful.

Against the remote Hindsight service at http://w4ce.jmsu.top:8888, retain/reprocess operations complete, but extraction currently produces:

• memory_unit_count=0
• total_nodes=0
• usage.total_tokens=0
• empty recall results

This indicates the CyberVerse payload is reaching Hindsight correctly, but the Hindsight backend extraction/LLM worker is not generating memory units yet. That needs follow-up on the Hindsight service side.

Tests

Executed:

git diff --check
pixi run -e macos-arm pytest tests/unit/test_hindsight_memory.py tests/unit/test_persona_agent_plugin.py -q
pixi run -e macos-arm go test ./internal/orchestrator
pixi run -e macos-arm python - <<'PY'
import yaml
from pathlib import Path
for path in ["infra/cyberverse_config.example.yaml", "deploy/gpu/cyberverse_config.gpu.yaml"]:
    yaml.safe_load(Path(path).read_text())
PY

Results:

• Python focused tests: 26 passed
• Go orchestrator tests: passed
• YAML parse: passed
• git diff --check: passed