Searchable data layer over Claude Code session transcripts

[DSado88]

The Insight

Claude Code generates rich .jsonl session transcripts at ~/.claude/projects/. This is the most complete record of AI-assisted software engineering that exists — every tool call, every file read, every decision, every mistake, every learning. But it's completely opaque:

• 374 project directories, 9.4 GB of session data on one developer's machine
• Single sessions can be 99 MB / 21,000 lines (21k structured events)
• Message types: assistant (8.5k), user (5.9k), progress (6.2k), file-history-snapshot (441), system (242)
• No search, no index, no way for Claude to learn from its own history

Claude has no persistent memory. But the transcripts ARE the memory — they're just not queryable.

Vision

A data layer (probably DuckDB-backed) that indexes Claude Code session transcripts so Claude can search its own history across sessions and projects. This may be a separate product from Squall, or a major extension of it.

What it could answer

• "What did we learn about auth last Tuesday?"
• "Find the session where we fixed the race condition in the review executor"
• "Which reviews found real bugs vs false positives?"
• "Show me every time I used tokio::select! — what patterns worked?"
• "What files do I modify most often across all projects?"
• "Summarize the 3 sessions I had on the memory module this week"

What it would index

From the jsonl events:

Field	Source	Value
session_id	Filename (UUID)	Groups all events in one conversation
project_path	Directory name	Which codebase
timestamp	Event field	When
type	assistant, user, system, progress, file-history-snapshot	What kind of event
tool_name	Tool use events	Read, Edit, Write, Bash, Grep, Glob, etc.
tool_input	Tool use parameters	File paths read/edited, commands run, search queries
content	Assistant/user text	The actual conversation
files_touched	Extracted from tool calls	Which files were read/modified
git_branch	Event metadata	Branch context

Derived/computed fields

Field	How
session_summary	LLM-generated summary of session purpose and outcome
decisions_made	Extract key decisions from assistant messages
bugs_found	Cross-reference with Squall review results
patterns_learned	Extract from memorize calls and skill updates
files_by_frequency	Aggregate file access across sessions

Architecture Sketch

~/.claude/projects/**/*.jsonl
        │
        ▼
   [ Indexer ]  ← runs on-demand or watches for changes
        │
        ▼
   ~/.claude/index/sessions.duckdb
        │
        ├── sessions(session_id, project, start_ts, end_ts, summary, git_branch)
        ├── events(event_id, session_id, ts, type, tool_name, content_preview)
        ├── file_access(session_id, file_path, access_type, count)
        └── search_index(event_id, content_fts)  ← DuckDB full-text search
        │
        ▼
   [ MCP Tool: "recall" ]
        │
        ├── recall search "race condition fix"
        ├── recall session <uuid> --summarize
        ├── recall files "src/memory/global.rs" --history
        └── recall week --summary

Open Questions

1. Separate product or Squall extension? Squall is a model dispatch + memory server. This is a session intelligence layer. Different concerns, maybe different binary.
2. Indexing strategy? Full ingest (9.4 GB) vs incremental (watch for new sessions) vs on-demand (index when queried).
3. Privacy/security? Session transcripts contain everything — API keys, file contents, passwords. Index must be local-only, encrypted at rest?
4. LLM summarization? Generating session summaries requires sending content to an LLM. Use local models? Squall's existing dispatch?
5. Scale? 374 projects × N sessions × 21k events per session. DuckDB handles this fine, but FTS index size?
6. Integration with Claude Code? Could this be an MCP server that Claude Code auto-connects to? Or a hook that runs post-session?

Why This Matters

Every Claude Code user generates this data. Nobody can search it. The developer who can query "what did I learn last week across all my projects" has a compounding advantage over the one starting fresh every session.

This is the missing feedback loop: sessions → index → recall → better sessions.

Data Sample

Current session (this Squall project):

21,492 lines / 99 MB
8,494 assistant messages
5,855 user messages
6,181 progress events
441 file snapshots
242 system messages

Full corpus:

374 project directories
~9.4 GB total
47 sessions in Squall alone