pi-memnest

English | 한국어

Connects pi to a running memnest service: five memory tools and task cards. The memory engine is not in this package, so start the core first.

Install

Needs Node.js 20+, pi, and a memnest service (normally http://127.0.0.1:3111).

pi install npm:pi-memnest@0.6.0

If Memnest is already registered from a local path or another version, remove that entry first to avoid duplicate tools. A pinned version is not changed by pi update --extensions.

Set MEMNEST_URL for another address and MEMNEST_TOKEN if the core requires a token. Check the connection with /memnest: it shows service health, memory count, data directory, whether Autocontext is on, the active task cards, and which card this session records into.

Tools

memory_remember text="Project X deploys on port 8320" memory_kind="fact"
memory_search query="Project X deployment port"
memory_remember text="Project X now deploys on port 8420" memory_kind="fact" supersedes="manual_..."

Without project, the extension sends pi's working directory and the core maps it to a private workspace. Ordinary memories are not encrypted at rest; enable the vault tools only for an agent you trust with plaintext credentials. The first search downloads the embedding model, so it takes longer.

Task cards

A task card is a handoff note for work that continues over several days. It works without any setting.

task_card action=save title="blog migration" text="# Goal ..."   # same title replaces the previous card
task_card action=list text="images"                              # rank by topic; omit text for newest
task_card action=get title="blog migration"                      # open and pin it for this session
task_card action=close title="blog migration"                    # stop recording when you move on
task_card action=done title="blog migration"                     # finish; save again to reopen

The first prompt of a session lists up to 20 active cards. An opened card is pinned into the system prompt, and every message you send is appended to it word for word, so a summary cannot drop what you said. Short one-time reminders cover the cases where models missed a card in real use: no card yet, a compaction, memory_remember with no card open, and an open card unsaved for 8 messages. Cards are shared with Claude Code, Codex, and other MCP clients through the core. How it works, measurements, and limits

Commands

These run in the pi terminal UI without the model, and nothing they show is sent to a model.

• /memnest: service health, memory count, task cards, and which card this session records into.

• /memnest-cards: list active task cards (newest first, [open] marks the ones this session records into) and finished ones; pick a card to read it, open or close it in this session, or mark it done (asks first).

• /memnest-search [project:<name>|project:all] <question>: the same search as memory_search, up to 10 results with project, date, type, score and an excerpt; pick one to read it in full. Without project: it searches the current folder's workspace; project: completes project names. With no question it asks for one.

• /memnest-recent [project:<name>|project:all]: the 20 newest memories saved by hand or by the model (not transcript turns), with project, date and an excerpt; pick one to read it in full.

• /memnest-review [project:<name>|project:all]: goes through saved rules and preferences, longest unconfirmed first. For each one you see the full text, its project, when it was saved and last reviewed, and choose Keep (records today as the review date, so it comes back last next time), Delete (moves it to the trash; restorable for 30 days), Skip, or Stop. Nothing is deleted unless you choose Delete for that memory; a summary is printed at the end.

The last two need a core with POST /recent and POST /reviewed (0.6.0 or newer); an older core answers with a one-line "too old" message.

Without a UI (pi -p) the commands print text and do not ask anything. If the core is not running they say so in one line.

Autocontext

Off by default. To add related memories to each prompt, set MEMNEST_AUTOCONTEXT_MODE=balanced before starting pi. The core decides which saved memories share the prompt's topic, so a prompt such as "ok, go on" gets nothing. Each card shows memory IDs so the agent can check the full text with memory_get; retrieved text is treated as reference data, not instructions.

Conversation capture

The extension stores no conversation itself. memnest watch on the core side captures pi, Claude Code, and Codex transcripts, so nothing is stored twice. See conversation capture and history import.

Development

npm install      # also builds dist/index.mjs, the file pi loads
npm run smoke
npm run e2e      # needs a core release build

Changelog, third-party notices

License

MIT © Blue-B