mnemo-replay

Mnemo Replay

Raw access to the source data behind claude-mnemo. This is a skill, not an MCP tool. Use normal file and shell tools to inspect the JSONL transcript and the SQLite database directly.

Use this only after recall has already narrowed you to a specific session, turn, or observation. mnemo-replay is for raw reads, not broad search.

Two data sources

Source	Contents	How to read it
JSONL transcript	Exact Claude session bytes: user prompts, assistant responses, tool calls, tool results, summary markers	Read, Grep
SQLite database	Indexed mirror: sessions, turns, observations, memories	sqlite3 in read-only mode

Use JSONL for exact wording and full payloads. Use SQLite for joined lookups and aggregates.

JSONL transcripts

Where they live

Claude Code stores transcripts under:

~/.claude/projects/<encoded-project-path>/<content-session-id>.jsonl

The encoded project path replaces / with -. Example:

/Users/alice/code/my-app
→ ~/.claude/projects/-Users-alice-code-my-app/<uuid>.jsonl

For mnemo worker sessions, the project path is ~/.claude-mnemo, so those transcripts live under the dedicated pseudo-project directory for that path.

The easiest handoff is from:

recall(id="S12", depth="expanded")

Expanded session output includes a raw: line with the absolute JSONL path.

Useful patterns

Open the transcript directly

Read(<absolute jsonl path>)

Find user-message boundaries

Grep(pattern='"type":"user"', path='<jsonl path>')

Find tool calls for one tool

Grep(pattern='"name":"Bash"', path='<jsonl path>')

Find compact boundaries

Grep(pattern='"type":"summary"', path='<jsonl path>')

Summary lines mark compact boundaries in the raw transcript.

SQLite database

Where it lives

Default path:

~/.claude-mnemo/claude-mnemo.db

If the install uses a custom data dir, adjust accordingly.

Read-only usage

sqlite3 ~/.claude-mnemo/claude-mnemo.db -cmd ".mode column" -cmd ".headers on" "<query>"

Do not write through sqlite3. Writes still go through remember.

Relevant tables

Table	Row =	Useful columns
sessions	One Claude Code conversation	id, content_session_id, project, title, content, insight, next_steps, last_compact_turn
turns	One user prompt in a session	id, session_id, prompt_number, status, title, content, type, tags, files_read, files_modified, tool_call_count
observations	One tool call in a turn	id, turn_id, tool_name, tool_input, tool_result, status, title, content
memories	Durable cross-session knowledge	id, type, scope, title, content, reasoning, application, tags, status

Useful queries

Map session id to transcript metadata

SELECT project, content_session_id
FROM sessions
WHERE id = 12;

List turns in one session

SELECT prompt_number, type, title, status
FROM turns
WHERE session_id = 12
ORDER BY prompt_number;

Recent turns that touched a file

SELECT s.id AS session, t.prompt_number AS turn, t.title
FROM turns t
JOIN sessions s ON s.id = t.session_id
WHERE (t.files_read LIKE '%src/auth.ts%' OR t.files_modified LIKE '%src/auth.ts%')
ORDER BY t.created_at_epoch DESC
LIMIT 20;

Recent active memories by tag

SELECT id, type, scope, title
FROM memories
WHERE tags LIKE '%"feedback"%'
  AND status = 'active'
ORDER BY COALESCE(updated_at_epoch, created_at_epoch) DESC
LIMIT 20;

When to use which source

Situation	Use
Exact user prompt wording	JSONL
Full, untruncated tool output	JSONL
Pre/post-compact transcript reconstruction	JSONL
Aggregates or cross-session joins	SQLite
Tag / type / file filtering	SQLite
High-level search or browsing	recall, not this skill

Guidance

• Always narrow with recall first.
• Use the raw: line from expanded session output as the copy-paste handoff.
• Treat SQLite as read-only unless you are going through remember.
• If you need exact bytes, prefer the JSONL over the indexed mirror.