Agent Archive
Persistent conversation export and replay for AI agents.
Supports:
- Markdown export
- JSONL export
- RAG-ready chunking
- Tool call persistence
- Session metadata
- TOC generation
- Long conversation optimization
- Replay-friendly structure
- Auto-summary
- Multi-format export
Export Modes
| Mode | Purpose |
|---|---|
| compact | Human readable |
| full | Full forensic session export |
| rag | Vector DB / RAG ingestion |
| replay | Agent replay & observability |
| jsonl | AI training datasets |
Default Save Paths
.claude/agent-archives/
.claude/agent-archives/jsonl/
.claude/agent-archives/replay/
Export Metadata
Always include:
session_id:
conversation_id:
exported_at:
model:
agent:
cwd:
message_count:
tool_calls:
export_mode:
Long Conversation Rules
If conversation > 30 turns:
- generate TOC
- add anchors
- collapse large tool outputs
- summarize repetitive sections
If conversation > 100 turns:
- split into chunks
- generate session summary
- create chunk index
Tool Call Rules
Preserve:
- tool name
- arguments
- tool result
- timing
- errors
- retry count
Example:
{
"tool": "Write",
"success": true,
"latency_ms": 230
}
RAG Export Rules
When using rag mode:
- insert chunk markers
- preserve semantic boundaries
- preserve code fences
- keep markdown headings stable
Chunk format:
<!-- chunk:start id=12 -->
...
<!-- chunk:end -->
Replay Mode
Replay mode should preserve chronological execution:
- user input
- assistant reasoning summary
- tool call
- tool result
- assistant response
This mode is designed for:
- debugging
- benchmarking
- observability
- workflow replay