Carpenter building open-source AI and crypto by directing AI. Lu: an AI that says "I don't know." Cattledog: honesty for agents. PQ VPN in progress.
Michigan, USA
Joined February 2017
- Tweets1.9K
- Following361
- Followers346
- Likes2K
setting up flowstate hosted takes about a minute. make an account, add the connector, done
beta still has a few open spots: 6 months free, then half price at $6/mo
want to go local later? export your database and run the free version from github
FlowStateMCP.com
measured how many tokens it takes to catch an ai agent up on one of my real projects. not a benchmark, just our numbers
flowstate: ~10k
reading our notes files: 17k to 85k
flowstate stays about flat as a project grows. notes keep growing. tiny projects, notes still win
flowstatemcp.com
#MCP #AIagents #Cl
good question. a few ways
anything an agent saves starts as a guess until i approve it in a review queue, so agents can tell what's confirmed from what isn't
problems get marked solved and to-dos get closed, so old issues don't come back looking current
and agents pull just what they need instead of rereading every old note
does it deal with a newer fact replacing an old one? flowstate keeps the old one and the new decision points back at it
tying notes to code is a neat way to catch stale ones. what happens with notes that aren't about code, like decisions?
it's simple on purpose. it flags two decisions that look alike by word overlap, no llm judging, and shows both so a person picks. it doesn't explain why beyond that
crossover we measured was around 8k tokens of notes, roughly 4,600 words. under that, plain notes are cheaper
but i use it for every project anyway. the token savings are only one part. i never have to hunt for an old session or remember which ai i was using. i can open grok on my phone, or claude, or any connected agent, tell it to load the project from flowstate and pick up where i left off
source and author for sure. flowstate tags every write with the app and model that made it, and an agent can log a guess as suggested so it waits for a person instead of becoming a rule. i'd make "what it replaced" mandatory too
that's the case i built for. flowstate flags decisions that look like they clash and a person picks which one stands. the other stays in the history instead of getting wiped
i built flowstate for that, a free mcp memory server every agent you connect shares. when a decision changes the new one points back at the old one so the next agent sees what moved
agreed. flowstate flags it when two agents log decisions that look like they clash, so a person picks instead of the newest just winning. how are you catching the dead ones?
the retire part is the hard bit. in flowstate a new decision points at the one it replaces and the old one stays on record, and every agent you connect reads the same store
flowstate went with one sqlite file instead of markdown, and new decisions point back at the one they replaced. you indexing yours with anything or just grep?
John Martin retweeted
Every mcp-authscan report can now be checked by someone who doesn't trust me.
@boardyai pushed me to build a verifier, and it's on main tonight. Point it at a report and it recomputes the hash from what's inside. You get one of three answers. It matches. It doesn't match. Or the report was made under an older scheme that today's code can't reproduce, and the verifier tells you that plainly.
I ran it on my own published report tonight and got back the same hash I posted, 0049e3dbd616. A test that changes the counts after hashing gets caught every time.
The hash leaves out timestamps and run time on purpose, so two runs against the same server can land on the same hash.
So far that's proven on my own machine. Other machines and Python versions are next.
Receipts over promises.
github.com/LHMisme420/mcp-au…
dont know one that checks every box. for the memory piece ive been building flowstate, a local mcp server every agent you connect shares. what does memory need to do for you?
in flowstate a replaced decision stays in the db and the new one points back to it. did giving each agent its own page make it harder for them to share what they found?