Git-native project memory for AI coding agents, with deterministic local recall, evidence gates, and layered delivery that quotes, points, or abstains.

version
0.8.2
GitHub stars
427

Install

/plugin install ownmem@ownmem

Run it inside Claude Code after adding the marketplace with /plugin marketplace add grpcer/ownmem.

Contents

Discovered from source

Skills 3

Open OwnMem Console

OwnMem Console renders this repository's memory metrics — the known false-delivery residual and the delivery funnel beneath it, recall latency, corpus health, governance — in the browser. It binds 127.0.0.1 only; nothing is served off this machine.

If npx ownmem dashboard fails because the layer is missing, the repository was initialized without it. Re-run init with --layers dashboard (see the init skill), then retry.

Start it for the user

The server stays resident. Start it in the background; if an instance is already running, the same command returns "reused": true instead of starting a second one.

Do not redirect --json output to /tmp or any other shared directory. That JSON contains a long-lived access token, and /tmp is commonly world-readable.

In Claude Code, use the Bash tool with run_in_background: true. In other hosts, start the process in the background without capturing the token to a world-readable file:

npx ownmem dashboard --json

Then read the URL from the private instance record (mode 0600 under .local-test/):

npx ownmem dashboard --status --json

Open the url field in the browser. Keep the #t= fragment — the access token lives in the URL fragment and never reaches the server; without it the page only shows an access notice.

The CLI defaults to a random port. Pass --port 45300 (or another free local port) when the user needs a stable bookmark, and keep using that port later.

After opening, run npx ownmem report --since 7d once and give the user a one-line summary: the north star (how much of what was reported as a wrong delivery is still answered with prose), latency P50/P95, and the most notable gap or warning. When the sample is small, say so instead of dressing process metrics up as adoption.

Lifecycle

npx ownmem dashboard --status --json
npx ownmem dashboard --stop
npx ownmem dashboard --open

Set up OwnMem in this repository

OwnMem requires Node.js 20.6 or newer. Install the reviewed dependency, then generate the host adapters:

npm install --save-dev ownmem
npx ownmem init --hook --hosts claude,codex

Adjust the flags before running:

  • --hosts: any comma-separated combination of claude, codex, cursor, gemini, grok, and generic — match the tools the team uses. The list is recorded, and running init again with a new list is how a host is added or removed. Grok CLI reads Claude Code's files, so name it alongside claude when both are used.
  • --locale auto writes the generated instructions in the system language; the default is English.
  • Drop --hook if the team does not want the recall hooks. --hook also installs the layers those hooks need, up to the local Web console; --layers core,gates,compiler,dashboard installs a subset instead.

Initialization creates .ownmem/ plus bounded ownmem-generated blocks in the host instructions, and preserves every line outside those boundaries. Verify:

npx ownmem init --check
npx ownmem audit

The npm package is the engine. This plugin is an optional shortcut and does not auto-update on every host. After a version bump (for example a Dependabot pull request):

npx ownmem init --update && npx ownmem init --check && npx ownmem audit

OwnMem recall and daily use

OwnMem is this repository's local, deterministic memory. Recall runs offline — no model calls, no network, no query-time token cost — and every memory is plain Markdown inside the repository.

If the repository has no .ownmem/ directory, OwnMem is not installed here; use the init skill from this plugin first.

Recall before changing code

npx ownmem recall -- "problem, path, or symbol"

For natural-language symptoms, pass two or three phrasings in one call and let the deterministic fusion rank them:

npx ownmem recall --multi -- "symptom in plain words" "root-cause term" "english symbol"

Recall answers in one of three tiers, and reading them as the same thing is the one mistake that undoes the design:

  • content — the memory's own text. It passed every gate; use it, then verify against live code, because a memory records what was true when it was written.
  • pointers — up to three titles with the command that opens each one. This is not a short answer. Open the topic before you rely on anything in it.
  • abstain — nothing is written down about this, and the envelope says which gate refused. Not a search failure to retry with different words.

Do not implement memory logic here

This skill only routes to the CLI. Flags, feedback names, and contracts belong to the installed package, not to this plugin. Host plugins are optional shortcuts and do not auto-update on every host, so a stale copy of this file must not teach a retired protocol.

Use npx ownmem --help and npx ownmem <command> --help for the current surface. After a version bump, update the repository package and run npx ownmem init --update.

Keep the memory healthy

npx ownmem audit
npx ownmem report --since 7d
npx ownmem dashboard --open

Write a new memory only for lessons that reading the code cannot recover — upstream quirks, toolchain traps, timing races. Keep entries small and let npx ownmem audit enforce the growth quota.

Sponsors

Details

Marketplace
grpcer/ownmem
Source
plugins/ownmem
Category
Community