groundwork grows a third axis: safety (guardrails) + quality (dev-loop) + continuity (memory-loop).
memory-loop (v0.1.0) — new plugin
A memory lifecycle layered on Claude Code's native file-based memory. Local-only, zero-config, MIT.
- Save gate — confirm tier (
long/short) and expiry with the user before anything is written; unverified guesses never enter long-term memory. - Tiered expiry — lapsed
tier: shortmemories are moved toarchived/at session start, never deleted; the expiry day itself stays live (exclusive boundary). - Safe by default — files without a
tierkey (including everything that existed before install) are never touched; uninstalling leaves all memories intact. - Learning-review nudge — every N responses (configurable) the agent is reminded to distill habits, skills, and memories from recent work.
- HABITS.md frame — turns corrections and incidents into standing behavior: positive practices (🟢) and hard lines (🛑), with merge-over-multiply and an escalation ladder to hooks/skills.
- Identity — a one-time, declinable offer to set names for the user and the assistant (the assistant may pick its own name), injected as context each session. No CLAUDE.md edits, ever.
- 23
batstests + shellcheck, wired into CI.
guardrails and dev-loop are unchanged.
Install
/plugin marketplace add choiyounggi/groundwork
/plugin install memory-loop@groundwork
Then ask for "set up memory-loop" to run the guided setup.
MIT.