@krzyzanowskim/pi-optmem0.2.2

pi-optmem

OptMem as a Pi extension. OptMem is Victor Taelin's permanent memory for AI agents; this package ships a TypeScript port, so installing the extension is the only setup. No python3, no curl | sh. Node builtins only.

Install

pi install npm:@krzyzanowskim/pi-optmem
pi install git:github.com/krzyzanowskim/pi-optmem   # from git; a local checkout: pi install ./pi-optmem
pi -e npm:@krzyzanowskim/pi-optmem                  # try once without installing

What it does

At session start it runs memo wake and puts your memory into the system prompt of every run, so it survives compaction. The snapshot stays fixed for the session (so the prompt cache holds) and is re-read after compaction or with /optmem; notes taken mid-session are in the transcript meanwhile. It registers a memo tool the agent uses instead of bash: wake, note, nap, recall, zoom, forget, config, import. Each call is drawn as one line (noted #41, compressed 0-15); expand it for the full output.

In the interactive TUI, the extension also works between runs, with a nested model call:

• Auto-note: after each settled exchange it may save one more memory line. The log is append-only, so each one shows in the status line (optmem: noted #412 "…") and in /optmem.
• Auto-compress: it does the compressions that notes make due, up to 32 per idle stretch, pausing when you start a run. The agent is no longer interrupted for them, unless the model call fails; then the agent does them until the next attempt succeeds. In print and RPC modes the agent does them, as upstream intends.

Settings

Environment variables set the defaults; /optmem switches last for the session. A small, cheap model is enough for both jobs.

/optmem alone shows the store, its size, the switches and this session's auto-notes, and re-reads the memory. /optmem recall <regex> searches it. /optmem config WAKE_LINES=300 grows the session-start read; 96 lines is about 8k tokens, and every line goes into every prompt.

The store is created on first use at ~/.optmem/memory, or $MEMORY_DIR if set.

The port

memo.ts is a TypeScript port of upstream OptMem (commit f5b6a24a59b4; memo is unchanged through upstream main at 1fb164c) with the same fixed-width record format, so both implementations can read one store. Differences:

• Locking: the Python original uses fcntl.flock (msvcrt on Windows); Node has no flock, so this port locks with an atomic mkdir. The two locks don't interlock, so don't run the Python tool at the same time as a Pi session on the same store. Sequential use is fine.
• recall takes JavaScript regex: no Python-only syntax such as (?i) or (?P<name>…). Searches are case-insensitive either way.
• Messages name the tool memo rather than the script's path, and init prints no AGENTS.md block, because the extension injects the instructions itself.

Standalone: node memo.ts wake. That needs Node 22.18+ or 23.6+ to run TypeScript directly (22.6–22.17: add --experimental-strip-types).

Don't edit anything under the memory directory; memo manages it. Parallel Pi sessions share one memory; subagents must not run memo (the injected instructions already say so).

License

Upstream OptMem has no license, so none is granted here for memo.ts, the port; it is included with attribution to Victor Taelin. The rest of this package (index.ts) is MIT.