mason

[mason @notmason__]

ai coding assistants have no memory. every session starts from zero.

so they'll happily re-do something you already tried months ago and threw out — because they can't see why you threw it out.

selvedge v0.3.7 gives your ai a memory of your project's decisions.

[mason @notmason__]

it's open source, runs entirely on your machine, and installs in one line:

pip install selvedge==0.3.7

the feature's called prior_attempts. more on what shipped:
https://t.co/c19HtPBUve

[mason @notmason__]

ai coding assistants have no memory. every session starts from zero.

so they'll happily re-do something you already tried months ago and threw out — because they can't see why you threw it out.

selvedge v0.3.7 gives your ai a memory of your project's decisions.

[mason @notmason__]

one design choice i'm proud of:

it stays quiet unless it's confident. if selvedge isn't sure the history is relevant, it says nothing rather than guess.

a wrong answer teaches your ai to ignore it forever. silence is recoverable. so the default is cautious on purpose.

[mason @notmason__]

craftsmanship is knowing how to work, art is knowing when to stop.

[mason @notmason__]

https://t.co/bDXXrccLJd

[mason @notmason__]

local SQLite, MIT, no telemetry.

pip install selvedge
https://t.co/lfHQHbly9q

longer post on what's changed since launch:
https://t.co/sJmRWhbz4O

would love to hear what breaks.

[mason @notmason__]

shipped selvedge v0.3.4 over the weekend. the install used to be six manual steps. now it's one command.

selvedge setup detects which AI tools you have (claude code, cursor, copilot) and wires all of them up in one pass.

thread on what changed.

[mason @notmason__]

also new: selvedge watch. live tail of new events as the agent works. polls under WAL mode so it never blocks the writer.

useful for the first 5 minutes after install when you actually want to see the wiring is working.

[mason @notmason__]

selvedge v0.2 → v0.3.2 in three days, all of it the unsexy stuff: concurrency safety, schema migrations, structured logging, and now v0.3.2 - a self-diagnostic command that tells you which step of your install is silently broken before your agent does.

selvedge doctor.

[mason @notmason__]

pip install selvedge

[mason @notmason__]

wrote up the full version on https://t.co/fHtV8JcCfM for anyone who wants the long-form: https://t.co/XFe38s9D8k

[mason @notmason__]

every codebase being written by ai right now has the same ticking time bomb:

the agent knows exactly why it changed each line. session ends. commit says "update schema." reasoning is gone forever.

i built selvedge to fix it

[mason @notmason__]

the bet: in 5 years every team will need an audit trail for ai-written code. compliance, onboarding, debugging, understanding why the code is the way it is

that infrastructure doesn't exist yet. selvedge is my first pass

mit, python, https://t.co/wZLrgfahzt

[mason @notmason__]

what ended up mattering most:

- changeset grouping so you can pull a whole feature back later
- git hook auto-links events to commits
- import tool backfills schema history from existing migrations
- all local sqlite, no cloud