Last week I gave my LLM a memory layer I call Cortex. This week I started feeding it something stranger: what I was listening to while I worked. A work session is not just the files you touched and the decisions you made. It has a texture. The track that was playing when something finally clicked is part of that memory, even if you would never think to write it down. So instead of throwing that signal away, Claude and I built a small observer to catch it. ———————————————————————— What it does, in three layers: 1. Passive. A tiny watcher checks the local music app once a minute and logs track, artist, and timestamp to a plain markdown file. No browser audio, no streaming history scraped, just what is actually playing on the machine. 2. Bookmarks. When a session opens or closes it drops a marker, so the log has boundaries instead of one flat stream of songs. 3. Flags. When a track lands on a moment that matters, I star it with one line of context. "This was playing when the gallery finally rendered." That markdown file is just another source the brain reads. Same rule as everything else: > Files own the truth. The brain owns the connections. ———————————————————————— Here is the part I do not know yet, and why it is interesting. The episodic layer now carries an ambient track. Does that change retrieval? When I come back to a problem, will the brain surface the session by its soundtrack the way a smell drags back one specific afternoon? Or is it just noise in the index? I genuinely cannot tell you. It has been running for two days. That is the honest bit. This is an experiment, not a feature. I built the observer in an afternoon because the cost of being wrong is a markdown file I can delete. The cost of being right is a brain that remembers the way humans actually do, by association and atmosphere, not just by fact. The try-me is attached if you want to point one at your own memory layer. It is about forty lines. Watch what your brain does with it before you decide whether it earned its place.

It's the end of the week and I had tokens to burn. So I left Claude running a self-improvement loop this afternoon and went out to color. Oh yeah, I've got into coloring lately. Most people working with AI are babysitting it. One prompt, wait, watch, next prompt. Stuck in the loop. Every output needs you, right now, watching. But the work isn't the prompting. The work is the workflow. ————————————————————————————————— Build it once and the shape changes: - In the loop: you design the process, set the gates, define what good looks like. - On the loop: it runs in the background. You check at the gates and redirect when needed. - Out of the loop: it runs without you. You start another Claude session, or you go and color. The loop I left running today didn't need me. I'd already done the part that needed me. That's the whole trick. Pay the attention up front, once, on the system. Then the system buys it back. You free up the time for another session, or for a life. ——————————————————————————————————— You don't scale by prompting faster. You scale by being in the loop once and building workflows. /A<3

Every LLM session boots blind. Close the window and it forgets everything: the notes, the transcripts, the decisions you already made. It all piles up write-only, never read back. So you re-explain yourself from scratch. Every single time. Bigger context windows do not fix this. They are rented, and they reset. So I stopped prompting harder and built a brain underneath the model instead. Cortex is a memory layer between your files and any model. The principle: Files own the truth. The brain owns the connections. ————————————————————————————————— How it works, borrowed from human neuroanatomy: 1. Regions. Every memory is tagged by where it lives: language, vision, episodic events, project structure, identity. Retrieval comes back grouped that way, not as one flat pile. 2. Tracts. Memories are wired by typed edges. Some are fixed (a prompt to its answer, an image to its caption). Others grow by use: anything retrieved together often enough forms a permanent link. Fire together, wire together. Cold edges decay. 3. Consolidation. The brain sleeps. Raw capture distils through five tiers (raw, session, daily, weekly, promoted fact). A year of thinking compresses about 37x, and every fact still walks back to its source. Nothing is deleted. Importance decays, rows do not. 4. Modes. The same brain retrieves differently in creative versus deep versus fast-recall. Neuromodulation as a config flag. 5. Auto-recall. A session-start hook injects the right context into turn one, and edits re-index on the fly. The brain reminds you. You never have to remember to ask. ————————————————————————————————— It runs locally, needs no API key, and stays model-agnostic. Claude, Codex, Gemini, or a model on your own machine plug in through injected context, a thin adapter, or MCP. Delete the index and rebuild it from your files with one command. Lose the index, lose nothing. Full deep dive and explainer at: https://www.aris-space.com/documents/memory/cortex

Creative reference libraries rot. You save things, you forget them. Your aesthetic lives between your eyes and your past work, and re-explaining it every brief doesn't scale. I run two streams. Trend cards distil external references: one card, one thesis sentence, operative moves. DNA profiles map my own work: every claim cites a real anchor example. Different questions. Don't mash them. Nothing publishes without passing a HALT gate. No card without a thesis I'd defend. No DNA claim without an anchor I can point to. Read the deep-dive: https://www.aris-space.com/documents/tools-and-plugins/lens Taste is a structured artefact, not a vibe. The gate is the discipline. //A<3

If switching LLM means rebuilding your stack, the LLM was never the asset. The thing you had to rebuild was. ———————————— The mistake I kept making I was building "Claude tools". Things that only worked when Claude was on the other end. The day I wanted the same capability in Codex or Kimi, I had to rewrite. Each port drifted. The version I built second was always slightly behind the first. The model isn't the centre of your stack. The contracts you give the model are. ———————————— The pattern that fixes it Three rules. All boring on purpose. 1. Markdown is the contract. A capability is a document the model reads, not a Python wrapper around an SDK. 2. Files are the memory. Shared markdown on disk at a stable path. Every harness reads the same files. 3. CLIs are the tools. Anything deterministic (search, similarity, transforms) is a shell command. Models shell out. No vendor SDK at the surface. No model-specific prompts. No proprietary format. Just files and shell. ———————————— What it buys The model becomes a knob, not a foundation. Hard reasoning task? Reach for the biggest model. Mechanical refactor? Cheapest fast one. Hundred-thousand-line transcript? Whichever has the longest context window today. Same capabilities every time. No port. No drift. The one rule that does the most work Build the interface, not the integration. SDKs are integrations. Markdown contracts, files, and CLIs are interfaces. When in doubt, lean interface. ———————————— Takeaway The model is the replaceable part. Your contracts are the asset. If your toolkit can't survive a model swap, you don't have a toolkit. You have a vendor relationship. ———————————— Full long-read with three diagrams (lock-in vs portable, the three-layer stack, the model-as-knob): → https://www.aris-space.com/documents/tools-and-plugins/the-model-is-the-replaceable-part //A<3