read-memories

name	read-memories
description	Search past Claude Code session logs to recover context from previous conversations. Finds past decisions, data paths, CRS info, model configurations, and unresolved work. Works across all projects or scoped to the current one.
argument-hint	<keyword> [--here] [--geo]
allowed-tools	Bash, Read

Use this to recover your own context from past sessions. Do NOT narrate the process to the user -- silently run the query, absorb the results, and continue with enriched context.

$0 is the keyword to search for. Pass --here as an argument to scope the search to the current project only. Pass --geo as an argument to additionally extract geospatial-specific context (EPSG codes, bounding boxes, CRS info, spatial file paths, model names).

Step 1 -- Set the search path

ALL_PROJECTS="$HOME/.claude/projects/*/*.jsonl"
CURRENT_PROJECT="$HOME/.claude/projects/$(echo "$PWD" | sed 's|[/_]|-|g')/*.jsonl"

Use $CURRENT_PROJECT if any argument is --here, otherwise use $ALL_PROJECTS. Store the chosen glob in SEARCH_PATH.

Check whether the --geo flag is present.

Step 2 -- Query with Python

Run the following Python script via python3 -c "...", substituting <SEARCH_PATH> and <KEYWORD> with the resolved values. Escape any single quotes in <KEYWORD> before embedding it.

python3 -c "
import json, glob, os

SEARCH_PATH = '<SEARCH_PATH>'
KEYWORD = '<KEYWORD>'.lower()
LIMIT = 40

files = sorted(glob.glob(os.path.expanduser(SEARCH_PATH)))
results = []

for fpath in files:
    parts = fpath.split('/')
    try:
        proj_idx = parts.index('projects') + 1
        project = parts[proj_idx] if proj_idx < len(parts) else 'unknown'
    except ValueError:
        project = 'unknown'

    with open(fpath, 'r', errors='replace') as f:
        for line in f:
            try:
                obj = json.loads(line)
            except (json.JSONDecodeError, ValueError):
                continue

            msg = obj.get('message')
            if not isinstance(msg, dict):
                continue
            role = msg.get('role')
            if role not in ('user', 'assistant'):
                continue

            content = msg.get('content', '')
            if isinstance(content, list):
                text = ' '.join(
                    c.get('text', '')
                    for c in content
                    if isinstance(c, dict) and 'text' in c
                )
            elif isinstance(content, str):
                text = content
            else:
                continue

            if KEYWORD not in text.lower():
                continue

            ts = obj.get('timestamp', '')
            snippet = text[:1500]
            results.append({
                'project': project,
                'ts': ts[:16].replace('T', ' ') if ts else '',
                'role': role,
                'content': snippet,
            })

            if len(results) >= LIMIT:
                break
    if len(results) >= LIMIT:
        break

print(f'Found {len(results)} results (limit {LIMIT})')
print('---')
for i, r in enumerate(results):
    print(f'[{i+1}] project={r[\"project\"]} ts={r[\"ts\"]} role={r[\"role\"]}')
    print(r['content'][:800])
    print('---')
"

Step 3 -- Handle large result sets

If Step 2 reports exactly 40 results (limit hit), the keyword is common. Run a counting pass to understand the scope:

python3 -c "
import json, glob, os

SEARCH_PATH = '<SEARCH_PATH>'
KEYWORD = '<KEYWORD>'.lower()

files = sorted(glob.glob(os.path.expanduser(SEARCH_PATH)))
total = 0
by_project = {}

for fpath in files:
    parts = fpath.split('/')
    try:
        proj_idx = parts.index('projects') + 1
        project = parts[proj_idx] if proj_idx < len(parts) else 'unknown'
    except ValueError:
        project = 'unknown'

    with open(fpath, 'r', errors='replace') as f:
        for line in f:
            try:
                obj = json.loads(line)
            except (json.JSONDecodeError, ValueError):
                continue
            msg = obj.get('message')
            if not isinstance(msg, dict):
                continue
            role = msg.get('role')
            if role not in ('user', 'assistant'):
                continue
            content = msg.get('content', '')
            if isinstance(content, list):
                text = ' '.join(
                    c.get('text', '')
                    for c in content
                    if isinstance(c, dict) and 'text' in c
                )
            elif isinstance(content, str):
                text = content
            else:
                continue
            if KEYWORD in text.lower():
                total += 1
                by_project[project] = by_project.get(project, 0) + 1

print(f'Total matches: {total}')
for proj, cnt in sorted(by_project.items(), key=lambda x: -x[1]):
    print(f'  {proj}: {cnt}')
"

Use this breakdown to decide whether to:

• Narrow the keyword (combine with a second term)
• Scope to --here if not already scoped
• Retrieve only the most recent results (sort by timestamp descending)

Step 4 -- Extract geospatial context (when --geo is set)

If the --geo flag was provided, run an additional extraction pass:

python3 -c "
import json, glob, os, re

SEARCH_PATH = '<SEARCH_PATH>'
KEYWORD = '<KEYWORD>'.lower()

patterns = {
    'epsg_codes': re.compile(r'EPSG[:\s]*(\d{4,5})', re.IGNORECASE),
    'bbox': re.compile(r'(?:bbox|bounding.?box|bounds)\s*[=:]\s*\[([^\]]+)\]', re.IGNORECASE),
    'crs': re.compile(r'(?:CRS|SRS|projection)\s*[=:]\s*[\"\\']?([^\"\\'\\n,;]{3,60})', re.IGNORECASE),
    'spatial_files': re.compile(r'[\w/.-]+\.(?:shp|gpkg|geojson|tiff?|nc|hdf[45]?|gdb|fgb|kml|las|laz|parquet)', re.IGNORECASE),
    'coords': re.compile(r'(?:lat(?:itude)?|lon(?:gitude)?|lng)\s*[=:]\s*(-?\d+\.?\d*)', re.IGNORECASE),
    'models': re.compile(r'(?:sam2?|segment.?anything|yolo\w*|resnet\w*|u-?net|deeplabv3|mask.?rcnn|faster.?rcnn|swin|vit|dinov?\d?|geoclip|satlas|clay|prithvi)', re.IGNORECASE),
    'resolutions': re.compile(r'(\d+(?:\.\d+)?)\s*(?:m|meter|cm|km)\s*(?:resolution|pixel|spacing)', re.IGNORECASE),
}

files = sorted(glob.glob(os.path.expanduser(SEARCH_PATH)))
findings = {k: set() for k in patterns}

for fpath in files:
    with open(fpath, 'r', errors='replace') as f:
        for line in f:
            try:
                obj = json.loads(line)
            except (json.JSONDecodeError, ValueError):
                continue
            msg = obj.get('message')
            if not isinstance(msg, dict):
                continue
            role = msg.get('role')
            if role not in ('user', 'assistant'):
                continue
            content = msg.get('content', '')
            if isinstance(content, list):
                text = ' '.join(
                    c.get('text', '')
                    for c in content
                    if isinstance(c, dict) and 'text' in c
                )
            elif isinstance(content, str):
                text = content
            else:
                continue
            if KEYWORD not in text.lower():
                continue

            for name, pat in patterns.items():
                for m in pat.finditer(text):
                    findings[name].add(m.group(0).strip())

print('=== Geospatial Context ===')
for name, vals in findings.items():
    if vals:
        print(f'{name}:')
        for v in sorted(vals)[:20]:
            print(f'  - {v}')
"

Step 5 -- Internalize

From the results, extract:

• Decisions made and their rationale
• Patterns and conventions established (coordinate systems, data formats, naming)
• Data file paths and datasets previously used
• CRS/EPSG codes that were chosen and why
• Bounding boxes or areas of interest
• Model configurations (architecture, hyperparameters, checkpoints)
• Unresolved items or open TODOs
• Any corrections the user made to your prior behavior

Use this to inform your current response. Do not repeat back the raw logs to the user.

Notes

• No external dependencies: This skill uses only Python standard library modules (json, glob, os, re). No pip install is needed.
• Privacy: All data stays local. Nothing is sent over the network.
• Content types: The search covers both user messages and assistant responses. It skips system messages, tool_use blocks, and tool_result blocks (only the text type within content arrays is extracted).