memory

name	memory
description	Ingest a Claude Code or pi-agent session into the llm-wiki knowledge base. Parses JSONL session files, extracts clean conversation transcript, drops it into raw/sessions/ for standard wiki ingest. Use when user says "/memory", "save session", "ingest session", or "remember this session".

Memory — Session Ingest for LLM Wiki

Parse an LLM coding session into a clean markdown document and drop it into the wiki's raw/sessions/ directory for standard ingest.

When to Use

• User says /memory, "save this session", "ingest session", "remember this"
• User wants to capture decisions, insights, or knowledge from a session into their wiki

Prerequisites

• The llm-wiki project must exist (check for CLAUDE.md with wiki schema)
• The wiki root is determined by: explicit user path > current working directory > ~/llm-wiki

Workflow

Step 1: Identify the Session

Determine which session to ingest. Options (try in order):

1. User provides a path — use that JSONL file directly
2. User says "this session" or "current" — find the current session:
   • Pi sessions: ~/.pi/agent/sessions/<encoded-cwd>/ — pick latest .jsonl
   • Claude Code sessions: ~/.claude/projects/<encoded-cwd>/ — pick latest .jsonl
3. User says "last session" or "previous" — pick second-latest from above
4. Ambiguous — list recent sessions with dates/sizes, ask user to pick

Path encoding: Both pi and Claude Code encode the cwd path:

• Pi: -- prefix, / → -, e.g. --Users-skynet-myproject--
• Claude Code: - prefix, / → -, e.g. -Users-skynet-myproject

Step 2: Parse the Session

Run the parser script bundled with this skill:

python3 <skill-dir>/parse-session.py <session-file> [--output <output-path>] [--format <pi|claude>]

The parser:

• Auto-detects format (pi vs Claude Code) from first line
• Extracts user messages and assistant text responses
• Strips: tool calls/results, system prompts, ACM status, hooks, metadata, thinking blocks
• Preserves: code blocks shared in conversation, key outputs mentioned in assistant responses
• Adds YAML frontmatter: title, date, session_id, project, duration_estimate
• Groups conversation into numbered exchanges (Human/Assistant pairs)

If parser fails, fall back to manual parsing:

• Read the JSONL file
• For each line, parse JSON
• Keep entries where type is "user" or "assistant"/"message" with role "assistant"
• Extract text content, skip tool_use blocks
• Format as clean markdown

Step 3: Review with User

Before saving, show the user:

• Session date and estimated duration
• Number of exchanges extracted
• Top topics detected (scan for headers, key terms, project names)
• Approximate word count

Ask: "What should I emphasize? Any topics to skip?"

Step 4: Save to raw/

Save the cleaned transcript to:

<wiki-root>/raw/sessions/<date>-<session-name>.md

Where:

• <date> is ISO date from session timestamp (e.g., 2026-05-23)
• <session-name> is kebab-case derived from: user-provided name > session project > auto-generated from topics

Step 5: Trigger Ingest

Tell the user the file is ready and suggest running the standard wiki ingest:

"Session saved to raw/sessions/<filename>. Ready for wiki ingest — want me to process it now?"

If user confirms, follow the standard ingest workflow from CLAUDE.md:

1. Read the source completely
2. Discuss key takeaways with user
3. Create source summary in wiki/sources/
4. Create/update entity and concept pages
5. Flag contradictions
6. Update index.md, log.md, overview.md

For the source summary, use source_type: session and adapt sections:

• Key Takeaways — decisions made, insights discovered, problems solved
• Detailed Summary — topic-by-topic breakdown (not chronological)
• Notable Quotes — keep if there are genuinely quotable insights; otherwise rename to Key Decisions
• Questions Raised — open questions, unresolved issues, future work identified

Session Format Reference

Pi Agent JSONL

{"type":"session","version":3,"id":"<uuid>","timestamp":"...","cwd":"..."}
{"type":"message","id":"...","message":{"role":"user","content":[{"type":"text","text":"..."}]}}
{"type":"message","id":"...","message":{"role":"assistant","content":[{"type":"text","text":"..."}]}}
{"type":"tool_call","id":"...","tool":"bash","input":{"command":"..."}}
{"type":"tool_result","id":"...","content":"..."}

Claude Code JSONL

{"type":"user","message":{"role":"user","content":"..."},"timestamp":"..."}
{"type":"assistant","message":{"role":"assistant","content":[{"type":"text","text":"..."}]},"timestamp":"..."}
{"type":"tool_use","..."}
{"type":"tool_result","..."}

Noise Filtering Rules

Always strip:

• Tool calls and tool results (bash commands, file reads, edits)
• System prompts and permission mode entries
• ACM status tags (<context-status>, <pruned-manifest>)
• Hook events (SessionStart, etc.)
• Thinking blocks
• File history snapshots
• Model change / thinking level entries
• Usage/token metadata

Always keep:

• User messages (full text)
• Assistant text responses (the actual conversation)
• Code blocks that appear in assistant text (not tool outputs)

Optionally keep (if user requests verbose mode):

• Key tool outputs that represent important results
• Error messages that led to decisions