Overview

Instruction-only local memory/learning system: tiered storage in ~/self-improving/ (memory.md HOT, projects/domains/ WARM, archive/ COLD), corrections and self-reflection logging, promotion/demotion rules, heartbeat state, and non-destructive updates to workspace AGENTS.md/SOUL.md/HEARTBEAT.md. boundaries.md forbids storing secrets; optional consent-gated clawhub install proactivity.

Key Advantages

1.No install spec or code files — instruction-only

2.No credentials required; explicit boundaries against storing secrets

3.Tiered memory, conflict resolution, compaction and transparency

4.Kill-switch and export/wipe flows; configPaths match purpose

Use Cases

Log corrections and preferences to ~/self-improving/
Self-reflection after significant work; promote after 3x
Memory stats, export, forget; heartbeat integration
Non-destructive steering updates to AGENTS.md/SOUL.md/HEARTBEAT.md

Evaluation Scores

9.0

/ 10

Reliability

9.0

Functionality

9.2

Usability

8.8

Safety

9.0

Performance

8.8

Compatibility

9.0

Based on 1 evaluation · Latest: 3/18/2026

Download Trend

Loading...

Evaluation History (1)

9.0/103/18/2026darwin 25.3.0

▼

OS: darwin 25.3.0LLM: default

OpenClaw: Benign, high confidence. Requirements and instructions coherent with stated purpose: instruction-only local memory/learning, reads and writes ~/self-improving/, may non-destructively update workspace steering files; no credentials or exotic installs. Before enabling: confirm comfort with agent reading/writing ~/self-improving/ and editing AGENTS.md/SOUL.md/HEARTBEAT.md; back up those files; review boundaries.md; only allow optional Proactivity install after reviewing that skill; verify kill-switch and export/wipe; limit file permissions on ~/self-improving/ if desired.

Self-Improving Agent (Proactive + Self-Reflection)