Every AI coding session has the same quiet problem: when it ends, everything it learned ends with it. Close the window, and the agent that just spent an hour discovering exactly why your build was broken knows nothing about it tomorrow. You either re-explain everything from scratch, or you re-discover the same bug the hard way, again. This workspace runs two small tools specifically to stop that from happening.
The actual problem
An AI agent's working memory — its "context window" — only exists for the length of one conversation. That's fine for a single self-contained task. It falls apart the moment your work spans multiple sessions, which is most real work: a project that takes weeks, picked up and put down, with other people (or other sessions of the same agent) touching it in between.
Without something to bridge those sessions, you get one of two bad outcomes: you manually re-brief the agent every single time (tedious, and you'll forget details), or you let it start cold and watch it re-derive things you already know, sometimes making the same mistake it already made and fixed last week.
Two tools, two different kinds of memory
Beads is a lightweight issue tracker built specifically for this workflow, not a general project-management tool ported over. Its job is short-term, structured state: what's done, what's in progress, what's blocked and why, what's been deliberately deferred. A new session starts by asking it one question — what's the current state of things — and gets back a real backlog instead of a blank slate.
Mnemoria handles a different kind of memory: not tasks, but hard-won knowledge. The distinction matters. An issue tracker answers "what's left to do." Mnemoria answers "what did we already learn the hard way, so we never have to learn it again."
A concrete example, not an abstract one
While converting this site's theme to work with a newer version of Zola, a debugging session turned up something genuinely non-obvious: a UI component's rendered HTML would silently get mangled if its own internal logic left behind so much as one blank line in the output, because of an obscure interaction with how the markdown parser recognizes raw HTML blocks. That took real, patient trial and error to track down — comparing working examples against broken ones piece by piece.
That discovery got written to mnemoria the moment it was confirmed. Weeks later, when the same theme-conversion process needed repeating for a second, different theme, the agent picked up that exact note before touching a single template file — and hit the identical bug in the new theme within minutes, recognized it immediately, and fixed it without a single wasted debugging cycle. That's the entire value proposition in one example: work done once should never need doing twice, and a memory system is what makes that actually true instead of aspirational.
Why this is bigger than one project
The pattern generalizes past blogging or even past coding. Any long-running collaboration with an AI assistant benefits from separating "what needs doing right now" from "what have we permanently learned" — and treating both as things worth writing down somewhere durable, rather than trusting either a human's memory or a single conversation's context window to carry it forward indefinitely. The tools are small. The discipline of actually writing things down when you learn them is the part that matters.