vapid thorn
Tell your OpenClaw:
go to plur.ai and install memory
Or:
openclaw plugins install @plur-ai/claw
You correct your agent's coding style on Monday. On Tuesday, it remembers! You teach it your deployment process. Next week, it still knows. Every correction becomes permanent knowledge that surfaces automatically when relevant.
This is what a ContextEngine built for learning looks like - not just compression, but actual retention that improves through feedback.
Knowledge is stored as engrams - in cognitive science, the physical trace a memory leaves in the brain. In PLUR, they're small YAML files on your disk. An open format any tool can read. You can cat them, grep them, git commit them. plur.ai/spec
Here's where it gets interesting: with enough engrams, PLUR starts extracting meta-engrams (experimental). You correct coding style in one project, deployment habits in another - PLUR notices the shared pattern: "validate before committing." Knowledge learned in one domain, applied to another. Transfer of knowledge.
We tested. A learning Haiku with PLUR outperforms Opus without it - 2.6x better, 10x cheaper. 86.7% on LongMemEval. The bottleneck was never the model. plur.ai/benchmark
One more thing.
Shared memory and learning across tools. Correct something in OpenClaw - your Claude Code, Cursor, Hermes, and CLI all know it. Same files, same directory. Across devices: plur sync.
What one agent learns, every agent knows.
As a community, we should demand interoperability between our AIs - from Claude Code to OpenClaw to Hermes Agent to any other AI. Your knowledge shouldn't be locked to one tool, one cloud, or one memory solution. Engrams are our proposal for that open format.
plur.ai | github.com/plur-ai/plur
Open source. No cloud. No cost. 1,100+ engrams in production. Contributions welcome.
🧪 What did your OpenClaw learn that surprised you?
🐛 Bugs → GitHub issues or reply here:
marble haven
shell portal
thorn roost
