memscribe-cli

memscribe (CLI)

The memscribe binary — the daemon and the toolbox for Memscribe's deterministic, zero-LLM transcript capture. It wraps the workspace crates (memscribe-core, -adapters, -io, -sink) behind six subcommands.

cargo run -p memscribe-cli -- <command> [args]
# or, once installed:
memscribe <command> [args]

The pipeline is deterministic and never calls a model. By default the redaction pass is on, so secrets are stripped before anything is written. See the workspace ARCHITECTURE.md for the pipeline and memscribe.example.toml for the config surface.

---

Commands

watch — the steady-state capture daemon

Tail discovered transcripts (and serve the hook endpoint), preparing nodes to a sink as they arrive.

memscribe watch [--tools claude,codex,gemini] [--sink ndjson|sqlite|memdb] \
                [--out FILE|-] [--root DIR ...] [--once] [--config memscribe.toml]

Flag	Default	Meaning
--tools	every adapter	Comma-separated tool slugs to watch (SourceKind::parse values).
--sink	ndjson	Sink target: ndjson, sqlite, or memdb (memdb needs the memdb feature).
--out	- (stdout)	Where prepared nodes go (a file for ndjson/sqlite); - is stdout.
--root	$HOME	Directory root(s) to scan for transcripts; repeatable.
--once	off	Drain what already exists and exit, instead of tailing live.
--config	—	Path to a memscribe.toml (see memscribe.example.toml).

hook — the hook handler

Reads a hook payload from stdin, records it, and exits 0 immediately. It never blocks the agent and never invokes a model. Agents wire this as their hook command.

memscribe hook < payload.json

parse — one-shot parse a transcript to NDJSON

The workhorse for tests and debugging: run one transcript file through the adapter and the full pipeline, emitting prepared nodes as NDJSON on stdout.

memscribe parse <file> [--as TOOL] [--no-redact]

• --as TOOL forces a specific adapter (claude_code, codex, gemini, otel, cursor, windsurf, zed, vscode, copilot). Omit it to infer the tool from the path; if inference fails the command tells you to pass --as.
• --no-redact emits verbatim content (used by golden tests that assert on exact text). Redaction is on otherwise.

memscribe parse ~/.claude/projects/foo/session.jsonl --as claude_code

replay — re-run preparation over a historical session

Re-prepares a transcript file with the current pipeline (redaction on). Useful after an adapter or pipeline change to see the new node stream for an old session.

memscribe replay <file> [--as TOOL]

verify — the conformance smoke suite

Parses every fixture under fixtures/ and prints a per-tool CASES / OK / NODES table, exiting non-zero on any failure. This is the fast, shellable summary the daemon ships with; full cross-tool conformance and the §8.3 invariants live in the testkit (cargo test -p memscribe-testkit).

memscribe verify
memscribe verify --capture     # (planned) snapshot a live session into a new fixture

redact — preview the redaction pass

Reads a file and prints it with secrets replaced by [REDACTED:<label>], warning on stderr if anything was stripped. --no-content elides all text and keeps only structure.

memscribe redact session.jsonl
memscribe redact session.jsonl --no-content

---

Logging

Logs go to stderr (stdout is reserved for node output), filtered by the standard RUST_LOG env var. The default level is warn.

RUST_LOG=debug memscribe parse session.jsonl --as codex

---

Build features

The CLI builds every adapter by default. The MemDB sink is feature-gated in memscribe-sink and off by default; --sink memdb only does anything once that feature is compiled in.

cargo build -p memscribe-cli     # ndjson sink (default), all adapters

The CLI does not yet expose a passthrough feature for memscribe-sink/memdb, so enabling the MemDB sink is a build-config follow-up: add a memdb = ["memscribe-sink/memdb"] feature to crates/memscribe-cli/Cargo.toml and build with --features memdb. Until then, --sink ndjson (the default) and --sink sqlite are the available targets from the binary.