home

bowerbird is a small, calm memory layer for AI agents. it stores natural-language notes as plain markdown files on disk, indexes them with hybrid BM25 + dense vector search, and uses an LLM to keep them curated as the corpus grows. one Go binary, one sqlite file, no daemon required.

get started the bower model

why bowerbird

one binary

15 MB Go binary, no daemon required.

filesystem first

markdown files are the source of truth; sqlite is a rebuildable cache.

hybrid search

BM25 + dense vectors + reciprocal rank fusion, with optional LLM re-rank.

add-only

new contradicting information sits alongside the old; retrieval handles recency.

runs anywhere

linux, macos, and yes, android via termux.

quick taste

# tell it something
bower curate "the deploy pipeline uses github actions then flux on the cluster"

# ask later
bower query "how do we deploy"

# let it think
bower dream

the satin bowerbird collects blue feathers, petals, berries, and bits of thread, and arranges them carefully at the entrance of its bower. bowerbird does the same with your agent's memories.