Garden of Thoughts | SherryAnalytics | July 4, 2026
Sub-title: Same memory directory, different note-taker.
This week I opened a fresh Claude Code session and asked it to help me share a blog post on LinkedIn. Before I explained anything, it already knew my workflow: run the publishing pipeline first, check the Open Graph tags, pre-warm LinkedIn's cache with the Post Inspector, draft the commentary for my approval — and share manually through the composer, not through the API script we wrote together, because API posts don't render the image card.
I never re-explained any of that. It came from a note the assistant wrote to itself a week earlier.
Since Fable 5 came back last week, I have been running it on the same project where Sonnet kept notes before it. That accidental experiment — same project, same notebook, two different models — taught me more about AI memory than any product announcement.
Here is the part that surprised me most: there is no vector database. No embeddings, no retrieval pipeline, no memory SaaS.
Claude Code gives every project a memory directory under your home folder — outside the repo, so it can never be accidentally committed:
~/.claude/projects/<your-project>/memory/
├── MEMORY.md ← index, loaded at the start of every session
└── linkedin-posting-workflow.md ← one fact per file
Each memory is a small markdown file with a few lines of frontmatter: a name, a one-line description, and a type (is this about the user, a project decision, or a piece of feedback?). MEMORY.md is a plain index — one line per memory — and it is the only thing loaded into the model's context at session start. The detail files get read when they become relevant.
That's the whole architecture. Persistent AI memory, implemented as files a human can open, edit, and delete.
The directory is provided by Claude Code, the harness. Any model running inside it — Sonnet, Opus, Fable — gets the same directory and the same instructions. So when I switched models mid-project, the notebook carried over untouched.
What changed was the judgment about what goes in it.
The clearest example: early in the LinkedIn project, the assistant saved a note describing the workflow we had built — a Python script that posts through LinkedIn's API. A few days later, I changed my mind. The API posts couldn't render the branded image card, and managing a 60-day access token wasn't worth it for a personal blog. I told the assistant I preferred sharing manually through LinkedIn's composer.
Fable 5 didn't append a second, contradictory note. It rewrote the existing one — restructured it around my new decision, kept the hard-won gotchas (LinkedIn caches card images by URL; published posts freeze their card forever), and demoted the API script to a footnote: still exists, if automation is ever wanted.
That is curation, not accumulation. A notebook where every entry is appended eventually becomes noise; a notebook that gets rewritten as understanding improves stays useful. Anthropic's own guidance for Fable 5 says as much: give it a memory surface — even a plain markdown file — and it performs measurably better, because it has been trained to write learnings down, consult them later, and update notes rather than duplicate them.
Sonnet used the same notebook competently when prompted. Fable treats it like something it owns.
In my last post I wrote about Recursive Language Models and Alex Zhang's idea that AI capability can come from a better working environment, not just a bigger model — give the model a REPL, and it stops being a passive reader and becomes an analyst with tools.
Memory is the same thesis wearing different clothes. Nothing about my assistant's brain got bigger between sessions. It got a notebook, an index, and the habit of maintaining both. The capability I experienced this week — an assistant that remembers my decisions and the reasons behind them — came from scaffolding: markdown files and a convention for using them.
Signoff: The next leap may not be a model that knows more. It may be a model that takes better notes.
This post was written, in whole or in part, with assistance from Claude (Fable 5) — which also wrote the memory notes it describes.