session-recall

Total recall for Claude Code. Get back everything compaction took.

npx session-recall "the thing you lost"

The problem

When Claude Code compacts context, you get a vague summary. Details vanish: decisions, error solutions, specific commands that worked, corrections you gave.

But the raw transcripts are never deleted. Every session is saved as JSONL in ~/.claude/projects/, permanently. Compaction only affects the live context window. The data is always there, just buried in JSON soup.

session-recall parses those JSONL transcripts properly. It extracts human-readable messages, tool calls, errors, and patterns, so you find what you need in seconds.

Search any past session

Find exactly what was lost after compaction:

Analyze your sessions

--report finds retry loops, errors, user corrections, inflated self-scores, and generates CLAUDE.md rules to prevent the same mistakes:

Cross-session patterns

--all analyzes your last N sessions to find recurring problems:

$ session-recall --all 10

CROSS-SESSION SUMMARY (10 sessions)
  Total tool calls: 4351 | Total errors: 165

RECURRING RETRY PATTERNS
  Bash: retried in 7/10 sessions (avg 4.2x when it happens)

SELF-SCORING ACCURACY
  185 scores across 10 sessions (avg 7.7/10)
  20/185 (11%) had user issues after

RECURRING ERROR TYPES
  COMMAND_FAILED: in 8/10 sessions
  FILE_NOT_FOUND: in 6/10 sessions

Deep analysis via Gemini

--deep sends structured session data to Gemini for project-specific insights instead of generic advice:

$ session-recall --report --deep

DEEP ANALYSIS (via Gemini)

1. PROJECT CONTEXT: Building a video generation pipeline with Remotion.

2. CLAUDE.MD RULES:
   - When ElevenLabs returns 429, wait 30s before retry. Agent wasted 20 calls.
   - Always check ffmpeg output file exists before proceeding to next step.
   - User wants narration synced with visual transitions, not just content.

3. BIGGEST TIME WASTER: 47 minutes retrying a Bash command blocked by a
   pre-commit hook. Switch approach after first hook rejection.

All commands

# Search
session-recall "keyword"              # Find keyword in current session
session-recall "error" "deploy"       # AND search (both must match)
session-recall --recent 10            # Last 10 messages (no tool noise)
session-recall --decisions            # Find decision points
session-recall --tools "Edit"         # Search tool calls only
session-recall --list                 # List all sessions

# Pin a session (auto-namespaced per Claude process)
session-recall --pin-by "project-x"   # Pin session containing keyword
session-recall --unpin                # Remove pin

# Analyze
session-recall --report               # Errors, retries, corrections, rules
session-recall --report --deep        # + Gemini project-specific insights
session-recall --all                  # Cross-session patterns (last 10)
session-recall --all 20 --deep        # Cross-session + Gemini
session-recall --check-compaction     # Exit 0 if session was compacted, 1 otherwise

# Apply rules (interactive review, then append to CLAUDE.md / MEMORY.md)
session-recall --apply                # Template-based rules
session-recall --apply --deep         # + Gemini project-specific rules

Apply rules to CLAUDE.md

--apply generates rules, shows each one for review, and appends approved rules to your project's CLAUDE.md:

$ session-recall --apply --deep

Running Gemini deep analysis...
Project: Building session-recall CLI tool for Claude Code transcript recovery.

==================================================
REVIEW CLAUDE.MD RULES (5 items)
==================================================
  [y] approve  [n] skip  [e] edit  [a] approve all  [q] quit

  (1/5) After 2 failures with Bash, switch approach. Do not retry a third time.
  > y

  (2/5) When Gemini returns malformed JSON, strip markdown fences before parsing.
  > e
  new text> Strip ```json fences from Gemini responses before JSON.parse().
  ...

Appended 4 rules to /root/my-project/CLAUDE.md
Appended 2 entries to ~/.claude/projects/-root-my-project/memory/MEMORY.md

MCP server

Give Claude Code direct access to past session context. Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "session-recall": {
      "command": "npx",
      "args": ["-y", "session-recall", "--mcp"]
    }
  }
}

Claude Code gets 6 tools:

Tool	What it does
recall_search	Search past sessions by keyword
recall_recent	Get last N messages (no tool noise)
recall_report	Analyze session patterns and suggest rules
recall_apply	Append approved rule to CLAUDE.md or MEMORY.md
recall_decisions	Find decision points
recall_list	List available sessions

After compaction, Claude can call recall_search to recover lost context, then recall_report to suggest rules, and recall_apply (with your approval) to persist the lessons.

Auto-trigger after compaction

Add a PreToolUse hook to ~/.claude/settings.json so Claude gets notified after context compaction. The hook checks once per session and reminds Claude that MCP recovery tools are available.

1. Create the hook script at ~/.claude/hooks/post-compaction-recall.sh:

#!/usr/bin/env bash
# Must ALWAYS output {"decision": "approve"} on stdout.
trap 'echo "{\"decision\": \"approve\"}"' EXIT

# Walk process tree to find Claude's PID (session-stable flag key)
CLAUDE_PID="" PID=$$
for _ in 1 2 3 4 5 6 7 8 9 10; do
    PARENT=$(ps -o ppid= -p "$PID" 2>/dev/null | tr -d ' ') || break
    [ -z "$PARENT" ] || [ "$PARENT" = "1" ] || [ "$PARENT" = "0" ] && break
    PID="$PARENT"
    PNAME=$(ps -o comm= -p "$PID" 2>/dev/null | tr -d ' ') || continue
    case "$PNAME" in claude|claude-code|node) CLAUDE_PID="$PID"; break ;; esac
done

FLAG="/tmp/session-recall-post-compact-${CLAUDE_PID:-$$}"
[ -f "$FLAG" ] && exit 0
touch "$FLAG" 2>/dev/null || true

# Try PATH, then common install paths, then npx
RECALL=""
if command -v session-recall >/dev/null 2>&1; then
    RECALL="session-recall"
elif [ -x /usr/local/bin/session-recall ]; then
    RECALL="/usr/local/bin/session-recall"
elif command -v npx >/dev/null 2>&1; then
    RECALL="npx -y session-recall"
fi

if [ -n "$RECALL" ]; then
    $RECALL --check-compaction 2>/dev/null && \
        echo "Context was compacted. MCP tools: recall_search, recall_report, recall_recent" >&2
fi
exit 0

chmod +x ~/.claude/hooks/post-compaction-recall.sh

2. Register the hook in ~/.claude/settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash|Read|Edit|Write|Glob|Grep",
        "hooks": [
          {
            "type": "command",
            "command": "~/.claude/hooks/post-compaction-recall.sh",
            "timeout": 5
          }
        ]
      }
    ]
  }
}

--check-compaction exits 0 if the current session was compacted, 1 otherwise. The flag file in /tmp ensures the check runs only once per Claude process (cleaned on reboot).

Setup

npx session-recall --help

For deep analysis and MCP --deep mode, add a Gemini key:

mkdir -p ~/.config/session-recall
echo "your-gemini-key" > ~/.config/session-recall/gemini-key
chmod 600 ~/.config/session-recall/gemini-key

Or set GEMINI_API_KEY as an env var.

Part of Building Open

Open-source tools for Claude Code power users:

• session-recall - Recover context after compaction
• claude-wrapped - Your Claude Code year in review
• bouncer - AI quality audit for any work
• blast-radius - Impact analysis before code changes

License

MIT