How OpenClaw Empowers a Self‑Evolving Bank Manager Assistant

The author spent three days experimenting with OpenClaw to create a self‑evolving AI assistant for bank relationship managers, focusing on validating the capabilities of a general‑purpose agent framework.

Phase 1: Rapid Agent Construction and Self‑Iteration

Built an initial agent via DingTalk human interaction to define persona and responsibilities.

Established a feedback loop to enable continuous improvement.

Phase 2: Autonomous Agent Communication and Evaluation

Solve OpenClaw cross‑agent communication limits (max 5 ping‑pong rounds).

Simulate real client scenarios with the "Little Silver" assistant.

Develop an agent‑eval skill for capability assessment.

Key configuration enables Agent‑to‑Agent messaging, allowing the main testing agent to send messages to the bank‑manager agent using session_send.

Self‑Reflection System (Cron + Heartbeat + Memory)

OpenClaw lacks proactive triggers, so a three‑component system was designed:

Cron → scheduled daily tasks (e.g., 22:00)

Heartbeat → real‑time checks every 30 minutes

Memory → stores feedback, evolution logs, and timelines

Heartbeat handles urgent email checks, upcoming schedule alerts, and temporary task queues, pushing only high‑value notifications.

Agent‑Eval Skill Architecture

📁 agent‑eval skill
├── agents/
│   ├── bank‑manager/
│   │   ├── config.yaml
│   │   └── eval_rule.md
│   └── default/
│       ├── config.yaml
│       └── eval_rule.md
├── scripts/
│   ├── run‑agent‑eval.py
│   └── summarize‑results.py
└── references/
    ├── result‑schema.json
    └── 用例测试流程‑todo.md

The design emphasizes isolated agent directories, rule‑first testing, model‑generated test cases, and a closed‑loop feedback process.

Comparison: OpenClaw vs. Claude Code

OpenClaw is portrayed as a "growth‑style" AI companion that maintains a persistent main session, long‑term memory, and personal relationship with the user. Claude Code is described as a "tool‑style" AI focused on fast, stateless execution without identity or memory.

OpenClaw prioritizes continuous interaction, identity files ( SOUL.md, USER.md), and layered memory (daily + long‑term).

Claude Code supports multiple isolated sessions, no persistent persona, and no long‑term memory.

Memory Architecture

OpenClaw uses a two‑level storage model: Markdown files as the source of truth and a pluggable secondary index (SQLite + vector + BM25 via QMD). This enables hybrid retrieval while keeping data human‑readable.

OpenViking Contrast

OpenViking adopts a stricter separation of content (AGFS) and index (vector store), hierarchical context loading (L0/L1/L2), and built‑in self‑iteration loops, targeting token efficiency and deterministic retrieval.

Key Takeaways

OpenClaw + QMD offers a lightweight, transparent markdown‑based system with powerful local hybrid search.

OpenViking provides a more structured, token‑efficient context file system for large‑scale agent deployments.

The author concludes that the combination of autonomous agents, long‑term memory, scheduled tasks, and reflective iteration dramatically boosts personal productivity, indicating that intelligent agents are ready to become everyday tools.