What I just pushed
I just pushed memory-spark! It is a memory plugin for OpenClaw agents. The code is at https://github.com/kleinpanic/memory-spark. There is no version number yet.
What it does
Before each turn, auto-recall injects the memories that fit. After a conversation, auto-capture stores the facts. A file watcher indexes your workspace files and sessions, plus PDF, DOCX and audio. Search is hybrid: dense vectors plus BM25 in LanceDB, with a cross-encoder to rerank the results.
How it works
It is TypeScript on Node 22 or newer. The pipeline runs extract, chunk, NER, embed, then upsert to LanceDB. The models run as services on a DGX Spark:
- embeddings from nvidia/llama-embed-nemotron-8b, 4096 dimensions
- reranking with llama-nemotron-rerank-1b-v2
- NER, zero-shot, OCR and speech to text
The embed queue is serialized and retries on failure. A dimension lock makes the plugin refuse to start if the provider or the dimensions do not match what the index was built with.
Install
Clone it, then run npm ci and npm run build. Next, enable it in your OpenClaw config. You need OpenClaw and your own embedding endpoints.
The first commit
The first commit is named "initial scaffold (broken)". About 13 minutes later came a commit that says the full implementation compiles clean. Then the README, a LICENSE and CI went in.
What is not done
There are no tags and no named versions. It needs OpenClaw and Spark endpoints, so it is not a drop-in plugin.
What is next
Try it and tell me what breaks. I want to know.
P.S Thanks for reading. Consider subbing to the RSS feed.
Klein Panic