1# Self-Improving Agent
2 
3A self-improvement system that captures learning artifacts from skill experiences and proposes validated updates.
4 
5## Overview
6 
7This agent captures reusable evidence from skill interactions. It implements a feedback loop with memory artifacts, self-correction proposals, and evolution markers. Durable skill or code changes still require validation or explicit approval.
8 
9## Key Features
10 
11- **Multi-Memory Architecture**: Semantic + Episodic + Working memory
12- **Evidence-Gated Learning**: Captures reusable lessons from skill workflows
13- **Pattern Extraction**: Converts experiences into reusable patterns
14- **Self-Correction**: Fixes skill guidance when errors occur
15- **Self-Validation**: Periodically verifies skill accuracy
16- **Proposal Artifacts**: Writes proposed updates before durable skill changes
17- **Confidence Tracking**: Measures pattern reliability over time
18- **Human-in-the-Loop**: Collects feedback to validate improvements
19 
20## Memory System
21 
22Current Claude Code hook integration writes to:
23 
24```
25~/.claude/memory/
26├── semantic/       # Patterns, rules, best practices
27├── episodic/       # Specific experiences and episodes
28└── working/        # Current session context
29```
30 
31## How It Works
32 
33```
34Any Skill Completes
35        ↓
36Extract Experience → Identify Patterns → Write Proposals → Consolidate Memory
37        ↓                     ↓                  ↓              ↓
38   What happened?    What can we reuse?   Which proposals? Track metrics
39```
40 
41## Installation
42 
43```bash
44apb skills add ./skills/self-improving-agent --scope global --target all --link
45```
46 
47## Hooks (Optional)
48 
49Wire hooks to capture errors and session-end signals:
50 
51```json
52{
53  "hooks": {
54    "PreToolUse": [
55      {
56        "matcher": "Bash|Write|Edit",
57        "hooks": [
58          { "type": "command", "command": "bash ${SKILLS_DIR}/self-improving-agent/hooks/pre-tool.sh \"$TOOL_NAME\" \"$TOOL_INPUT\"" }
59        ]
60      }
61    ],
62    "PostToolUse": [
63      {
64        "matcher": "Bash",
65        "hooks": [
66          { "type": "command", "command": "bash ${SKILLS_DIR}/self-improving-agent/hooks/post-bash.sh \"$TOOL_OUTPUT\" \"$EXIT_CODE\"" }
67        ]
68      }
69    ],
70    "Stop": [
71      {
72        "matcher": "",
73        "hooks": [
74          { "type": "command", "command": "bash ${SKILLS_DIR}/self-improving-agent/hooks/session-end.sh" }
75        ]
76      }
77    ]
78  }
79}
80```
81 
82## Triggering
83 
84### Host-Supported Follow-up
85When the host runtime supports hook follow-ups, this skill can be recorded or run after high-signal workflows such as:
86- prd-planner
87- code-reviewer
88- debugger
89- refactoring-specialist
90- etc.
91 
92### Manual
93```
94"自我进化"
95"self-improve"
96"分析今天的经验"
97"总结这次教训"
98```
99 
100## Example Learning
101 
102### Episode
103```yaml
104Skill: debugger
105Situation: Form submission doesn't refresh data
106Root Cause: Empty callback function
107Pattern: Always verify callbacks have implementations
108Confidence: 0.95 → Proposals: debugger, prd-implementation-precheck
109```
110 
111### Skill Update
112```markdown
113## Proposed Update (2025-01-11)
114 
115### Pattern Added
116**Callback Verification**: Always verify that callback functions
117passed as props are not empty and actually execute logic.
118 
119**Source**: Episode ep-2025-01-11-003 (3 occurrences)
120**Action**: Propose adding to debugger checklist
121```
122 
123## Research Basis
124 
125- [SimpleMem: Efficient Lifelong Memory](https://arxiv.org/html/2601.02553v1)
126- [ACM Memory Mechanisms Survey](https://dl.acm.org/doi/10.1145/3748302)
127- [Lifelong Learning of LLM Agents](https://arxiv.org/html/2501.07278v1)
128 
129## Templates
130 
131Reusable templates live in `skills/self-improving-agent/templates`:
132- `pattern-template.md`
133- `correction-template.md`
134- `validation-template.md`
135 
136## License
137 
138MIT
139