Interlinked.

Memory · why agents forget

Why does my AI agent keep forgetting?

It is not a bug, and a bigger model will not fix it. An agent's working memory is its context window: a fixed budget of text it can hold at once. Within a task it fills up, and the oldest turns get squeezed out to make room. Between tasks it starts empty: every new session is a blank slate. The two fixes everyone reaches for (ChatGPT memory and a CLAUDE.md file) each patch one kind of forgetting and miss the rest. Here is the actual mechanism, and the piece all of them are missing.

The mechanism, drawn

Two ways to forget, one context window.

The window is what the model can "see" right now. It is finite, so long tasks overflow it; it is volatile, so it empties when the session ends. Draw three sessions on a timeline and the amnesia is obvious, and so is the fix.

TIMESESSION 1Monday · fresh startSESSION 2Tuesday · fresh startSESSION 3next week · fresh startLEARNED THIS SESSIONtokens expire in 35dCARRIED OVER ✓tokens expire in 35dwithout it: blank, re-derives MondayCARRIED OVER ✓tokens expire in 35d+ adds what it learns, for next timeRESET↺ window clearedRESET↺ window clearedA PERSISTENT LAYER ON YOUR FILESknowledge written on the file itself, not the session. It survives every reset abovewrites ↓↑ returns↑ returnscontext window: volatile, cleared every sessionpersistent layer: keyed to the file, outlives the session

Two forms of forgetting, one root cause. The window is finite, so on a long task the earliest context is pushed out to make room for the latest. And it is volatile, so when the session ends everything in it evaporates: the next session opens with none of it. Rules files and memory features exist to reload some of that context next time. But a file the agent has to remember to open, that you keep accurate by hand, is a patch on volatility, not a cure for it. What actually survives forgetting is a layer that lives on the files themselves and is handed back automatically, on the searches the agent was already going to run.

Where agent memory lives today

Two famous patches. Both help. Both stop short.

ChatGPT memory
vendor account memory
Remembers
facts about you, across chats
Lives in
your account, in the cloud
Read by
ChatGPT only
Blind to
your files, your code, your machine
CLAUDE.md / .cursorrules
a rules file you maintain
Remembers
what you typed in, by hand
Lives in
one path, in one repo
Read by
one agent, in that repo
Goes stale
on rename, or when you forget to update
A file-keyed vault
what Interlinked adds
Remembers
notes + relationships on every file
Lives in
your machine, keyed to file identity
Read by
every agent (19 set up at install)
Survives
renames, new sessions, new vendors

Neither is bad. Keep them both. ChatGPT memory is the right home for "I prefer TypeScript": it persists across sessions, but it never indexed your machine and only one vendor reads it. CLAUDE.md sits on your side of the fence (real project knowledge), but you maintain it by hand, it covers one repo, one agent reads it, and it goes stale the moment you reorganize (we took that specific failure apart in Your agent's memory dies on rename). Each patches one leak. None is machine-wide, shared across agents, and self-maintaining at the same time.

What survives forgetting

A memory that lives on the files, not in the window.

Interlinked keeps agent knowledge in a vault on your disk (short notes and file-to-file relationships) keyed to each file's OS identity, the permanent handle Windows assigns the moment a file is created. Because the key is the file and not a path, a note follows a rename or a move without noticing. Because the vault is machine-wide, it spans every repo plus the ~90% of the disk that was never in git. And because every agent set up at install can read and write it, what Claude Code learns on Monday, Cursor inherits on Tuesday.

It also starts full instead of empty. On a fresh install of the test machine below, the vault seeded itself from the git history already on disk, so the very first session already had context the window alone could never carry across the gap:

12 / 12
repos found and webbed on the first pass
8,455
git events replayed in 1.7 s
26,958
relationship pairs, from receipts (no model)

Measured on the test machine below: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4.47M files.

The principle

The window resets. The vault remembers.

A note an agent leaves is searchable in about a millisecond, and the whole persistent layer rests near 44 MB, small enough to forget it is running, permanent enough that your agents never start from zero again. It is not a replacement for your model's memory or a rules file; it is the layer underneath both, the one the context window structurally cannot be.

1 identity
per file, assigned by Windows, permanent across renames and moves; the key everything hangs on
19 clients
configured at install; each reads and writes the same vault; no vendor owns your memory
100% local
the vault is a sealed store on your disk; nothing leaves the machine, ever

THE MEMORY BELONGS TO YOUR MACHINE, NOT TO A MODEL, NOT TO A VENDOR.

Using it

Keep your memory files. Add the one they can't be.

1

Keep what works. Your CLAUDE.md and your ChatGPT memory still do their jobs: project instructions and personal preferences. Don't delete a thing.

2

Install once. One signed installer sets up 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest. The persistent layer starts the moment it is installed.

3

Let it seed. Day one it reads the git history already on your disk. From then on it learns as your agents work: notes and relationships accumulate on the files themselves, no upkeep.

4

Stop repeating yourself. What one agent learns, the next inherits, across sessions, across vendors, across every repo on the machine. The blank slate is gone.

Everything local is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows · Next: why memory dies on rename

Measured July 2026 · Ryzen 9 9950X3D · 4.47M files · Windows 11← All research