mnemo-replay

name: mnemo-replay description: Read a past turn's full content — the user prompt, the complete assistant narration, and every tool call's raw input/output — directly from the mnemo SQLite database, which stores them uncapped (recall and timeline only show a truncated slice). Use after recall/timeline has pointed you at a session or turn. The raw JSONL transcript is a fallback only for exact bytes or message metadata the database does not mirror.

Mnemo Replay

Raw, uncapped access to what happened in a past turn. This is a skill, not an MCP tool.

Use it only after recall or timeline has already narrowed you to a specific session or turn — mnemo-replay reads a known target, it is not for broad search.

Read from the SQLite database first. It holds the full text that recall and the memory agent only ever see truncated: the user prompt, the complete assistant narration, and every tool call's raw input/output — all uncapped. Reach for the raw JSONL transcript only for the few things the database does not mirror (exact bytes, message metadata).

SQLite database (primary)

Where it lives

~/.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

Reconstruct an entire turn

Prompt, complete assistant narration, and every tool call's raw input/output — entirely from the database, no JSONL:

SELECT user_prompt, assistant_transcript
FROM turns
WHERE session_id = 12 AND prompt_number = 3;

SELECT tool_name, tool_input, tool_result, status
FROM observations
WHERE turn_id = (SELECT id FROM turns WHERE session_id = 12 AND prompt_number = 3)
ORDER BY id;

• assistant_transcript is the full interleaved narration — every assistant text block of the turn. assistant_response is only the final block (what the extractor and the recall preview see). Populated for turns recorded since the column was added; older turns leave it NULL — fall back to the JSONL for those.
• observations.tool_input / tool_result are stored uncapped, even for skipped rows.

Other useful queries

Map session id to transcript metadata (also gives you the JSONL location if you need the fallback)

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;

JSONL transcript (fallback)

Go here only for what the database does not store: exact verbatim bytes, the precise text/tool interleaving order, thinking blocks, per-message usage/model metadata, compact summaries, or turns predating the assistant_transcript column.

Where it lives

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

The encoded project path replaces / with - (e.g. /Users/alice/code/my-app → ~/.claude/projects/-Users-alice-code-my-app/<uuid>.jsonl); mnemo worker sessions use the project path ~/.claude-mnemo. The exact absolute path is the raw: line in recall(id="S12", depth="expanded") output — copy-paste it.

Parse it with the bundled script

Prefer the parser over hand-written Read/Grep — it understands resumed transcripts, promptId turn boundaries, compact markers, tool results, and malformed-line skips the same way claude-mnemo does:

bun "$CLAUDE_PLUGIN_ROOT/scripts/replay-parse.cjs" schema <jsonl-path>           # available fields
bun "$CLAUDE_PLUGIN_ROOT/scripts/replay-parse.cjs" show <jsonl-path> T12          # one turn
bun "$CLAUDE_PLUGIN_ROOT/scripts/replay-parse.cjs" query <jsonl-path> -f "promptNumber,assistantText:160" --grep "auth race"

Use raw Read/Grep only when even the script's output is insufficient — inspecting bytes it truncated, verifying an exact JSON field it does not print, or checking transcript corruption by line number.

On compacted sessions the JSONL's T<n> (promptNumber) can drift from the database's prompt_number (the DB also counts compact-continuation and superseded forked turns), so the same T12 may land on a different turn. When in doubt, anchor by --grep content rather than the turn number.

Guidance

• Always narrow with recall / timeline first; mnemo-replay reads a known target.
• For a turn's full content — prompt, complete assistant narration, every tool call's raw I/O — query the database: turns.user_prompt + turns.assistant_transcript + observations.tool_input/tool_result, all uncapped. This is the default path.
• Treat SQLite as read-only; writes go through remember.
• Go to the JSONL only for what the DB does not store (exact bytes, message metadata, interleaving order, or NULL assistant_transcript on older turns), and prefer replay-parse.cjs over hand-parsing.