Memex Search

Memex Search

Use this skill to index local history and retrieve results in a structured, LLM-friendly way.

Indexing

• Build or update the index (incremental):
  • memex index
• Continuous index:
  • memex index-service enable --continuous
• Full rebuild (clears index):
  • memex reindex
• Embeddings are on by default.
• Disable embeddings:
  • memex index --no-embeddings
• Backfill embeddings only:
  • memex embed
• Common flags:
  • --source <path> for Claude logs
  • --include-agents to include agent transcripts
  • --codex/--no-codex to include or skip Codex logs
  • --model <minilm|bge|nomic|gemma|potion> to select embedding model
  • --root <path> to change data root (default: ~/.memex)

Search (LLM default JSON)

Run a search; output is JSON lines by default.

memex search "query" --limit 20

Each JSON line includes:

• doc_id, ts (ISO), session_id, project, role, source_path
• text (full record text)
• snippet (trimmed single-line summary)
• matches (offsets + before/after context)
• score (ranked score)

Mode decision table

Filters

• --project <name>
• --role <user|assistant|tool_use|tool_result>
• --tool <tool_name>
• --session <session_id> (search inside a transcript)
• --source claude|codex
• --since <iso|unix> / --until <iso|unix>
• --limit <n>
• --min-score <float>

Grouping / dedupe

• --top-n-per-session <n> (top n per session)
• --unique-session (same as top-k per session = 1)
• --sort score|ts (default score)

Output shape

• JSONL default (one JSON per line)
• --json-array for a single JSON array
• --fields score,ts,doc_id,session_id,snippet to reduce output
• -v/--verbose for human output

Background index service (macOS launchd)

memex index-service enable
memex index-service enable --continuous
memex index-service disable

Narrow first (fastest reducers)

1. Global search with --limit
2. Reduce with --project and --since/--until
3. Optionally --top-n-per-session or --unique-session
4. memex session <id> for full context

Practical narrowing tips

• Start with exact terms (quoted) before hybrid if results are noisy.
• Use --unique-session to collapse PR-link spam fast.
• Use --min-score to prune low-signal hits.
• Use --sort ts when you want a timeline view.
• Use --role assistant for narrative outcomes; --role tool_result for command errors.
• For a specific session, prefer search "<term>" --session <id> --sort ts --limit 50 to jump to outcomes.

Config

Create ~/.memex/config.toml (or <root>/config.toml if you use --root):

embeddings = true
auto_index_on_search = true
model = "potion"  # minilm, bge, nomic, gemma, potion
scan_cache_ttl = 3600  # seconds (default 1 hour)
index_service_mode = "interval"  # interval or continuous
index_service_interval = 3600  # seconds (ignored when mode = "continuous")
index_service_poll_interval = 30  # seconds

auto_index_on_search runs an incremental index update before each search. scan_cache_ttl sets the maximum scan staleness for auto-indexing. index-service reads config defaults (mode, interval, log paths). Flags override. Service logs and the plist live under ~/.memex by default.

Recommended when embeddings are on (especially non-potion models): run the background index service or index --watch, and consider setting auto_index_on_search = false to keep searches fast.

Semantic and Hybrid

• Semantic: --semantic
• Hybrid (BM25 + vectors, RRF): --hybrid
• Recency tuning:
  • --recency-weight <float>
  • --recency-half-life-days <float>

Fetch Full Context

• One record:
  • memex show <doc_id>
• Full transcript:
  • memex session <session_id>

Both commands return JSON by default.

Human Output

Use -v/--verbose for human-readable output:

• memex search "query" -v
• memex show <doc_id> -v
• memex session <session_id> -v

Recommended LLM Flow

1. memex search "query" --limit 20
2. Pick hits using matches or snippet
3. memex show <doc_id> or memex session <session_id>
4. Refine with --session, --role, or time filters