Features · pop culture · measured July 2026
Black Mirror imagined total recall. Your AI agent has amnesia instead.
In 2011, Black Mirror aired “The Entire History of You”, an episode where a small implant behind the ear, the grain, records everything you see and hear, so any moment can be replayed, searched, and zoomed into at will. It imagined perfect, total, instantly-searchable memory. It also spent the whole episode arguing that this was a curse. Today's AI agents have the exact opposite affliction: not too much memory. None. They forget between sessions and re-learn your whole machine every time you say hello.
TL;DR · the short version
The fiction imagined too much memory. “The Entire History of You” (Black Mirror, 2011) gave people perfect, replayable recall of their whole life, and treated it as horror.
Your agent has the opposite disease: amnesia. It forgets between sessions and re-reads your machine from scratch each time: on the test machine that cost 6m 57s and 71 tool calls to reach one file.
CLAUDE.md is the industry’s workaround, and it rots: notes are hand-written and chained to a path, so a folder rename orphans the memory.
The benign version keys memory to your files, not to you: to each file’s OS identity, so it survives renames, spans every repo, is read by every agent, and never leaves the machine.
It builds itself and stays cheap: on a fresh machine it turned 8,455 git events into 26,958 relationships in 1.7 seconds, and the always-warm engine idles near 44 MB, less than one browser tab.
First, the fiction (accurately)
The show dreamed the wrong dream.
“The Entire History of You” is the third episode of Black Mirror's first series (2011), written by Jesse Armstrong, later of Succession. Its premise is a brain implant, the grain, that gives everyone perfect, replayable, searchable memory of their entire life: rewind an argument, zoom in on a face, project the footage on the wall for someone else to watch. It is exactly the total recall a good assistant seems like it should have.
And the whole episode is about why that is a nightmare. The protagonist replays a loved one's memories over and over, hunting for betrayal, until the marriage collapses and he ends the story gouging the grain out of his own neck. The moral is not “perfect memory is good.” It is that a device recording a person (every glance, forever, un-deletable) is a trap. The dream was aimed at the wrong target. The thing you actually want remembered is not your face. It is your work: which file, which decision, what connects to what. Same idea, pointed at your files instead of your eyeballs, and suddenly it is useful instead of horrifying.
Three kinds of memory, side by side (the fiction, the tool you use today, and the benign inverse):
| Fiction’s grain | Today’s agent | Interlinked’s vault | |
|---|---|---|---|
| What it remembers | everything you see + hear | nothing: forgets each session | your files + how they connect |
| Keyed to | a person’s body | a path, repo, or account | the file’s OS identity |
| Survives a rename? | n/a | no: the note orphans | yes: never notices |
| Who can read it | you, and whoever you show | one vendor’s agent | every agent you connect |
| Leaves your machine? | it’s implanted in you | vendor memory does | never: offline by design |
| Who builds it | constant recording | you, hand-writing CLAUDE.md | itself, from your own history |
| Can you delete it? | only by surgery | scattered, per tool | yes: you own it |
The grain remembered the person and could never let go. The vault remembers the work and answers to you: every row you can verify, and every row you can delete.
Too much · too little · just right
The spectrum of memory, drawn honestly.
On the left, the fiction: total recall of a whole person, un-forgettable and un-deletable, a surveillance curse. In the middle, where your agent actually lives: no recall at all, re-learning the disk on every run. On the right, the benign middle nobody wrote a dystopia about: memory keyed to your files, built from history you already have, read by every agent, and yours to erase.
The trick was never “remember everything.” It was “remember the right thing, keyed to the right anchor.” Key it to a person and you get Black Mirror. Key it to the file's own identity and you get a layer that survives a rename without noticing, the same fix drawn out in your agent's memory dies on rename.
The cost of starting from zero
Amnesia isn't free. You pay for it three times.
Every new session, an agent wakes up knowing nothing about your machine and has to re-derive it. We watched Claude Code do exactly that (reach a single file it had no memory of) on the test machine below. A forgetful floor charges the same bill, over and over:
- 1
It re-learns the machine. 6 minutes 57 seconds and 71 tool calls to find one file (a coffee run and change) versus 16 milliseconds and a single call once the layer remembers where things live. That is roughly a 7,200,000× gap on one task, purely because the memory wasn’t there.
- 2
You pay to re-teach it. About 58% of a session’s tokens (more than half of a $200 agent plan) burned re-reading directory listings it already read yesterday and forgot. Amnesia isn’t just slow; it’s metered.
- 3
The workaround rots. So people hand-write CLAUDE.md, .cursorrules, and vendor-memory files to re-teach the agent by hand. But those are chained to a path or an account: rename a folder and the note orphans, silently.
None of this is the model's fault. It's a memory problem, one layer down. If you've ever wondered why your AI agent forgets, this is it: the working memory is the context window, and it clears when the session ends. The fix isn't a bigger brain. It's a memory that lives on the machine and outlasts the chat.
The benign version
Total recall, pointed at your files.
A knowledge vault is a machine-wide store of short notes and relationships, keyed to your files, built automatically and read by every agent. It is the grain's idea with the ethics flipped back to sane. It doesn't record you. It records your work: which files change together, which repos share code, what an agent learned about a file last week. Each note is keyed to the file's OS identity, not its path, so it survives renames and moves, spans every repo and the non-git 90% of the disk, and is read by whatever agent you happen to open.
And you don't feed it. On a fresh machine, the first time it ran, it read the deepest history already sitting there (the git log) and turned 8,455 events into 26,958 relationships in 1.7 seconds, across all 12 repos, before a single search was run. No model, no embeddings, no cloud. The opposite of the grain on every axis that mattered: it remembers the work, not the person; it lives on your disk, not in a vendor's; and you can delete any of it, because you own it.
It isn't a vector database and it isn't your model's memory feature. It's the layer under both. See how the files start finding each other on day one, and why it's not RAG and not memory.
THE GRAIN RECORDED THE PERSON. THE VAULT REMEMBERS THE WORK, AND YOU CAN DELETE IT.
Questions people ask
The short answers.
What did Black Mirror’s “The Entire History of You” actually show?
It is the third episode of Black Mirror’s first series (2011), written by Jesse Armstrong. In it, a small implant behind the ear (the grain) records everything a person sees and hears, so any moment can be replayed, searched, and zoomed into on demand. The episode is a cautionary tale: the protagonist replays memories until his marriage collapses and finally cuts the grain out of his own neck. This article uses it as a metaphor; it is not affiliated with or endorsed by the show or its makers.
Why do AI agents “forget” between sessions?
An agent’s working memory is its context window, which is cleared when the session ends. Nothing persistent sits on your machine describing where things are or what was decided, so the next session starts blank and re-reads the disk from scratch. That re-learning is the tax we measured: 6 minutes 57 seconds and 71 tool calls to reach one file on the test machine.
How is Interlinked’s vault different from CLAUDE.md or vendor memory?
CLAUDE.md and rules files are hand-written and chained to a path: rename the folder and the note orphans. Vendor memory is tied to your account and lives in that vendor’s cloud. The vault keys its notes and relationships to each file’s OS identity, so they survive renames and moves, span every repo and the non-git files too, are read by every agent you connect, and never leave the machine.
Does the vault record my screen or send anything to the cloud?
No. It is the opposite of the grain. It stores short notes and relationships derived from history you already have, such as your git commits, keyed to files. There is no model, no embeddings, no recording of you, and nothing is uploaded. It runs offline, and you can delete any of it because you own it.
How do I try it, and what does it cost?
One signed installer wires up 19 AI clients at once: Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest. Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo.
Give your agent a memory
Cure the amnesia. Skip the dystopia.
Install once. One signed installer builds the index and auto-configures 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest. No JSON, no keys.
Let it seed itself. The first run reads your git history and webs your repos together in seconds: 8,455 events into 26,958 relationships on the test machine, before you ask a thing.
Point your agents at it. Every connected agent now shares one machine-wide memory instead of re-learning the disk each session, and it stays near 44 MB at rest.
Rename freely. Move folders, rename files, reorganize. The notes are keyed to each file’s OS identity, so the memory follows: nothing to orphan, nothing to leave the machine.
Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows · Next: why your agent's memory dies on rename