memory-bridge: remember.sh hung silently instead of reporting a down core

[joelteply]

The memory tool stored nothing and said nothing — the worst combination for a system whose job is preserving corrections.

Two defects in one line:

if "$CONTINUUM" memory/remember ... 2>/dev/null | grep -q '"appended"...'

1. No timeout. memory/remember dispatches to a core; with none answering the CLI cold-starts one (a full Rust build via start-server.sh), so this blocked for minutes. Measured: a remember sat 300s and was killed by the caller (exit 143).
2. 2>/dev/null discarded the cause. The fallback could only guess ("is the server up?"), throwing away the CLI's own stderr — the one diagnostic that would have named it.

Together they made the engram path look permanently broken, and it was never investigated because the failure had no voice. An agent believes its correction is durable, drops it from context, and re-learns the same lesson next session — precisely what this plugin exists to prevent, happening inside the plugin.

Now

Bounded call (REMEMBER_TIMEOUT, default 20s), stderr captured and reported, and the outcomes named distinctly — a timeout says the core did not answer and NOTHING WAS STORED; a non-zero exit prints the CLI's real error. Both state "nothing was stored" explicitly, because silence is what let this survive.

Not the underlying cause

continuum start shells to a build script, so a cold core is minutes away and callers can't distinguish that from a hang. That belongs with core lifecycle work — an install-registered supervisor, and a start that launches the installed server binary instead of compiling.

bash -n clean.