cheesedb
Agent memory where deleting a source deletes what it taught
TLDR
The project calls this a preview, not a production store.Create an agent
$ cheesedb agent create [alice]
Add a source and print its id$ cheesedb lot [alice] [note] "[text]"
Add a memory that depends on source ids$ cheesedb remember [alice] "[text]" "[1 2]"
List every record$ cheesedb show [alice]
List live memories$ cheesedb active [alice]
Search raw files for text$ cheesedb find [alice] "[text]"
Search by meaning (needs an embedding model)$ cheesedb search [alice] "[question]" [10]
Ask a language model using those memories$ cheesedb ask [alice] "[question]"
Remove what a source taught$ cheesedb recall [alice] [id] fair erase
Check the log$ cheesedb verify [alice]
SYNOPSIS
cheesedb COMMAND [ARGS]
DESCRIPTION
cheesedb is a local memory store for agents. Data lives under ~/.cheesedb unless this variable points somewhere else:
$ CHEESEDB_DIR
The directory is created mode 0700, with data/, tokens/, and demo/. The Go module path is github.com/govinargo-labs/cheesedb/cmd/cheesedb. Go 1.26.5 or newer is required. Linux is tested. Windows is not supported.Commands from the program's usage text:- agent create / agent delete — an agent and its files. The token is stored under tokens/- lot AGENT TYPE TEXT — add a source and print its id- remember AGENT TEXT LEGS — add a memory. A leg is ids joined by commas (4,5 means both). A space-separated list (1 2) means either- show — every lot and memory- active — live memories- get AGENT ID — one record- find AGENT TEXT — search the files on disk- search AGENT QUESTION [K] — K nearest memories by embedding. Default 10- embed AGENT — make live memories searchable. Later remember calls do this themselves- index AGENT — list indexes. index AGENT drop NAME drops one- recall AGENT ID fair|strict untrusted|erase — drop a lot or memory and what depends on it- forget AGENT TEXT — find the closest memory, show what would go with it, erase after a yes- compact AGENT — erase the text of dead records- verify AGENT — check the log and print the head fingerprint- version — print the version- learn AGENT — turn live sources into memories with a chat model- ask AGENT QUESTION [K] — answer from the K closest live memories only- demo — a scripted comparison with a naive store- mcp AGENT — serve that agent's memory as MCP toolsEmbedding and chat use llm.conf in the home directory. The keys are:$ embed_base_url
$ embed_model
$ embed_api_key
$ base_url
$ api_key
CAVEATS
The author marks the tool as a preview. forget and recall ... erase delete text. agent delete removes that agent's files. An embedding server and a chat server are optional until you use search, embed, learn, ask, or forget.
