Explainer · ~10 min read
The complaint is universal: "I told it last week, why doesn't it remember?" The honest answer isn't a bug or a quota — it's that the thing people call AI memory is actually five different things in a trench coat, and none of them work the way human memory does. Once you can see the five layers, the forgetting stops being mysterious.
One-sentence answer: Out of the box, a language model is a stateless function — text in, text out, nothing carried over between calls.
When you send a prompt to GPT-5 or Claude or Gemini at the API level, the model sees exactly what you put in this request and nothing else. There's no hidden "session." If you want it to know what was said three turns ago, you have to include those three turns in this turn. The conversation feel of ChatGPT is the client stitching turns together and resending them — not the model remembering.
Once you accept that, every "memory" you see in a product is a layer added on top.
Read these in order. They're additive, not interchangeable.
The literal text the model sees on this turn. Hard-capped by the model: GPT-5 family runs 128K–400K tokens depending on tier, Claude Sonnet/Opus is 200K, Gemini 2.5 Pro is 1M. Once a conversation exceeds the cap, the client either truncates or summarizes older turns. The model has no idea it lost anything.
A small structured store the client re-injects into the system prompt on future chats. ChatGPT's memory is typically 1,200–2,000 tokens of pinned facts. Claude's "memory" is really Project context — you set it, it persists per Project. Gemini's is closer to a cross-Google-account context than discrete pinned facts. None of these store conversations; they store distillations.
An external index — usually a vector store — that the client searches at query time and injects relevant chunks into the context window. This is what "AI that knows your documents" actually means under the hood. It works well when retrieval is tuned; it fails silently when the right chunk isn't in the top-K.
Modifies the model's weights so it produces text in a particular style. It does not store facts you can read or edit. A fine-tuned model can sound like your past chats without containing them. Asking it "what did I tell you last week?" returns a confident hallucination, every time.
A static block of text the client prepends to every chat in this account or Project. Counts against the context window. Effective but tiny — usually 1,500 characters of soft cap, and identical for every conversation.
One-sentence answer: Human memory is retrieval over a lifetime of episodic detail; every layer above is either too small, too lossy, too static, or too inaccessible.
A human who has known you for a year doesn't recall every conversation, but they can usually answer "what was that thing you told me about your sister last fall?" That requires (a) a durable store, (b) cued retrieval, (c) reasonable confidence calibration, and (d) the ability to forget gracefully. The provider memory features do (a) partially, (b) almost not at all, (c) badly, and (d) by quietly evicting things they decided weren't important.
The 1,200–2,000 token ChatGPT memory cap, for context, is roughly 150 short facts. A year of meaningful conversation with another person produces tens of thousands of those.
When Gemini 2.5 Pro shipped at 1M tokens and Claude added 200K, a common reaction was: problem solved. It isn't. Three reasons:
Four properties — the missing layer the providers don't ship:
Provider memory features ship #3 partially and a thin slice of #4. Dedicated memory layers (Mem, Rewind in its old form, Konshus, hand-rolled RAG setups) try to do all four.
Konshus is the missing layer. Import your existing conversations from ChatGPT and Claude, drop in journals, docs, and meeting transcripts; the vault distills them into atoms — small facts with source, confidence, and a timestamp. You see every atom. You can edit any of them. When you talk to an AI, the vault hands it a tight context block tuned to the current question. Model updates and provider sunsets don't touch the vault. See the full backup guide for how to seed it from what you already have.
Konshus holds your context outside any single provider. Model updates, account changes, and quiet sunsets don't touch the vault. Encrypted, exportable, never used for training.
Meet Konshus