Your assistant forgets. Your coding agent starts from scratch. Switch apps and you explain yourself all over again. ctxv0 keeps what’s true about you right now — and hands it to whichever AI you’re talking to.
It updates. It cites. It declines.
Tell it things over time. Ask it anything. When what you said has changed, it gives you the current answer and keeps the old one. When your sources contradict each other, it tells you that instead of choosing.
from ctxv0 import Memory m = Memory(user="alex") m.remember("Launch is on the 12th.") m.remember("Launch moved to the 19th.") m.state("launch date") # → "the 19th" # the 12th: retired, kept with its date m.remember("Marco says the launch is the 26th.") m.state("launch date") # → known: False # reason: conflicting_sources # candidates: ["the 19th", "the 26th"]
Tell ChatGPT you went vegetarian and Claude has no idea. Teach your coding agent your conventions and the next tool starts blank. Every AI company keeps your history inside its own walls. ctxv0 puts it back in the middle, where all of them can reach it.
Most AI memory piles facts up and ranks them. Four things make this different, and none of them need explaining twice.
When something changes, the old version is retired with the date it stopped being true — not quietly left in the pile to resurface six months later.
“You changed jobs in March, so it stops introducing you by the old title.”
When two things you said contradict each other, it shows you both and asks, rather than ranking them and answering in a confident voice.
“You told me the 19th, then Marco said the 26th. Which one?”
Paste in an email, a transcript or a script and other people’s words don’t become facts about you. It records who said what.
“That was a line in the document you shared, not something you told me.”
Every fact points back to the exact sentence, the speaker and the date. Nothing in your memory is unexplained, and anything can be removed for good.
“I know that because of what you wrote on 14 March.”
Same memory underneath, four ways in.
Add memory to an agent you’re building. You set the token budget; it returns a small block of resolved facts instead of a wall of transcript.
block = m.context(query, budget_tokens=4000)
Connect it to Claude, Cursor, or anything that speaks MCP, and the same memory is live in all of them. No copying context between tools.
claude mcp add ctxv0 --url https://your-instance/mcp
Every user gets an isolated memory resolved from their credential alone. One person’s facts are never reachable from another person’s session.
A managed endpoint and a key, so there’s nothing to run yourself. Early-access users get it first and help decide what it should do.
The usual way to give a model memory is to paste the whole history back in. On a standard 500-question benchmark, that is both the expensive option and the less accurate one.
per 1,000 questions · 103,904 tokens each · 61.0% correct
per 1,000 questions · 8,464 tokens each · 71.0% correct
A twelfth of the cost and ten points more accurate, on the same questions with the same model. Figures are input tokens at published gpt-4o-mini rates.
| What the AI was given | Correct | Context used | |
|---|---|---|---|
| No memorythe question on its own | 9.8% | — | |
| The entire historyeverything, pasted in | 61.0% | 103,904 | |
| ctxv0just what matters now | 71.0% | 8,464 | |
| ctxv0the same run, repeated | 72.4% | 8,464 |
Benchmark LongMemEval, 500 questions · answering model gpt-4o-mini at temperature 0 · graded by the benchmark’s own unmodified scorer running on gpt-4o · every row identical except the context handed to the model.
A memory that stays correct as life changes is the same hard problem whether it serves one agent or a person moving between a dozen AI apps. Models will keep getting smarter. They still won’t know what changed about you last Tuesday.
The library and the MCP server, in private beta, with early users shaping what ships.
A managed endpoint and a key, so durable memory is a five-minute job rather than a project.
A personal memory you own and carry between AI apps — portable by design, so switching tools never means starting over.
Free while it’s in beta, no card, no waitlist queue. Tell me what you’re building or what you use AI for, and I’ll set you up and walk you through it myself.
Get early access