SLOPIE
IT ATE MY HISTORY. ALL OF IT.
IT REMEMBERS WHAT YOU SAID TO YOUR AGENTS!free · npx slopie audit
i work with two agents. in two different apps.
the api changed. use v2.
using v1!
(shouts) %@#!
neither remembers what i told the other. or me.
so i got bored and built slopie. it eats every session and remembers.
it counts how much i swear at them. 358 times. 1 in 23 prompts.
honestly? lower than i thought.
slopie
memory for coding agents
my agents had no memory across sessions, and codex and claude code never matched up. everything i’d told them was already on my disk. nobody was reading it.
anyone running more than one coding agent, at the moment one session hands work to the next.
solo: a typescript cli and a plugin per agent. github’s only other contributor is claude (12 commits). it used to be called potsherd; the tools still are.
- every agent session on your machine
- captured and indexed, locally
- mcp: recall → read → graft
- the next agent gets exact quotes, with their source
potsherd_recall · _read · _graft · _write
no model in the middle. a handoff is built from exact quotes, not summarised by an llm that might make things up. and only an explicit write becomes memory, so an agent can’t remember something by accident.
a fresh agent tried a release candidate cold, on a half-indexed archive. first recall: 77 seconds, and an irrelevant fragment. it gave slopie 4/10 and found the answer in the repo’s files instead. cold retrieval is the open problem.
limits: not certified as unattended primary memory. the audit may send redacted chunks to opencode zen and typesafe/jev for free analysis, with a notice first; POTSHERD_OFFLINE=1 keeps it local.
1,031 npm downloads, 27 sep–26 oct · v1.8.0 · node 22+ · mit
$ npx slopie auditall 28 of my last public pushes went here.
the end.
of issue #2 SLOPIE anyway.