eywa

name: eywa description: > Cross-session memory and continuity CLI. Provides session memory, handoff documents, and historical session retrieval. Use at session start to retrieve context from past sessions (eywa get). Use at session end to persist a handoff document for future sessions (eywa extract). Trigger on: "session memory", "handoff", "continuity", "past sessions", "what was I working on", "session start", "session end", "batch index", "historical sessions", "prior context", "previous session".

Eywa CLI

Cross-session memory. Extracts handoffs at session end, retrieves past context at session start.

Execution

All commands use this exact pattern — no variations:

cd "$EYWA_REPO" && python -m eywa.cli <command> [args]

Where $EYWA_REPO is the path to this repository (the directory containing the eywa/ package). If using a virtualenv, activate it first or use its Python binary directly.

Copy-paste commands

# Get recent sessions (default 3)
cd "$EYWA_REPO" && python -m eywa.cli get

# Get by keyword search
cd "$EYWA_REPO" && python -m eywa.cli get "mcp routing"

# Get with options
cd "$EYWA_REPO" && python -m eywa.cli get "topic" --days-back 30 --max 5

# Extract current session (auto-detect)
cd "$EYWA_REPO" && python -m eywa.cli extract

# Extract specific session (8-char short ID or full UUID)
cd "$EYWA_REPO" && python -m eywa.cli extract 1b2f6f6b

# Rebuild index
cd "$EYWA_REPO" && python -m eywa.cli rebuild-index

Output goes to stdout, errors to stderr. Exit code 0 = success, 1 = failure.

Why cd + python -m: The eywa package lives in the skill directory, not on PATH. The venv Python picks it up from CWD. Do not try bare eywa — it does not exist as a binary.

When to Call

eywa get: Session start (no query), when user asks about past work (keyword query), when you need prior decisions.

eywa extract: Session end, after significant milestones, when user asks to save context.

Extract is a two-step protocol:

1. Summary first. Before running eywa extract, produce a concise session summary inline — what was built, what decisions were made, what's open. This is your own synthesis, not a tool call. Write it as a message to the user.
2. Then extract. Run eywa extract (or eywa extract <session_id>). The CLI independently parses the JSONL transcript. Your summary above becomes part of the transcript, enriching the handoff with an explicit recap.

Summary format depends on channel:

Claude Code PC session (default):

• A table of items built/changed (columns: item, status, notes)
• Followed by 2-3 succinct paragraphs: decisions made, patterns discovered, open threads

Telegram session:

• Short bulleted list of items done
• Short bulleted list of pending/open items
• 2-3 succinct paragraphs covering decisions and context

eywa rebuild-index: After manually editing handoff files, corrupt index, after bulk imports.

Session Detection (extract)

Without session_id: auto-detects via PID tracing → CWD-scoped mtime → global mtime. Requires JSONL modified within 30 seconds.

With session_id: explicit lookup only (8-char short ID or full UUID).

In multi-session environments, pass session_id explicitly.

Bundled Resources

Tips

• Keywords, not sentences. "mcp server" not "let's continue working on the MCP server"
• Don't extract trivial sessions. Empty work = empty handoffs.
• max > 5 is capped. 3 is usually enough — more dilutes context.
• Check exit codes. Exit 1 means failure.

Anti-Patterns

• Do not run eywa extract on trivial sessions. Sessions under 3 turns or 400 chars of content produce low-value handoffs.
• Do not skip eywa-batch on first install. Without historical context, eywa get returns nothing useful.
• Do not use eywa get without a query when many handoffs exist. It returns recent sessions by default — use keywords for targeted retrieval.
• Do not hardcode paths. Use env vars (EYWA_DATA_DIR, EYWA_SESSIONS_DIR) for portability.