Your work is summarised remembered.

Native Context Memory keeps what you’ve said, done and taught in the model itself. Tell it once; it’s there in every session.

Memory and knowledge, kept with the model.

The model provides the intelligence. AdaptiqAI adds memory and your knowledge. Native Context Memory is built to work across open models.

Memory is held by the engine that serves the model. The weights stay unchanged, and your apps share the memory for the model you choose.

An exploded technical drawing. At the bottom sits the open-weight model. Two thin layers are part of it and lift out: Native Knowledge, holding your documents such as pricing.pdf and brand-voice.md, and above it Native Memory, holding what was done, such as “Launch: Friday 16 October”. Above the block, three of your apps share one line into the memory. One app writes the launch date; later another reads it back without anyone sending it again. When the layers sink back, the seams disappear and it is one block again.
  1. Your apps and harnesses

    All of them share one memory.

  2. Native Memory

    What you’ve done, kept between conversations.

  3. Native Knowledge

    Your documents — add, correct or remove them.

  4. Open-weight model

    The open model you choose.

    MistralMiMo · XiaomiNVIDIA · NemotronDeepSeekReflection · BeamQwenKimiGLM
UsuallyMemory sits beside the model and is read again every time you ask.
Native Context MemoryMemory is held by the engine that serves the model, not re-read from your prompt.

Today, nothing is remembered. It is re-read.

It feels as if your AI tools remember. They don’t: the whole conversation is sent again with every message and read from the start, until it no longer fits. That re-reading is one of the reasons token use gets so high.

Without memory

Every message sends the whole conversation again. When it reaches the limit, it is summarised, and the reasons go first.

With Native Context Memory

No more compacting. Only the new message is read. The conversation is already held in memory by the engine that serves the model, so nothing is summarised away.

Work the way you always do.

You write to the model the way you always have, in the harness you already use. Behind it, every moment is kept with its time, and the model recalls the one it needs — from your conversations or your knowledge — without pasting it into the prompt. And you have far more context to work with: up to 5M-token context in each agent session.

See the recall benchmark
  1. Kept with its time
  2. Recalled, not compacted
  3. Yours to review, edit or delete
  4. Add your own knowledge

Let your agent work for hours.

As the work grows, a 1M context window fills up. The context is compacted, the agent loses the thread and comes back to you. With up to 5M-token context in each agent session and no compacting, it keeps the thread from the first ticket to the last.

ExampleWork through 2,000 support tickets

Four things it is not.And what it does instead.

Techniques it is often mistaken for. Most of them send text along with every call, and finetuning makes a new model.

Not a context window

What it does

A context window is how much fits in one call. All of it is sent and processed every time, and when it is full, something is cut.

What Native Context Memory does instead

Native Context Memory is held by the model’s serving engine, not re-sent in the prompt. Nothing is summarised away; when it is full, older conversations stop being recalled first.

Not RAG

What it does

Retrieval fetches passages that look like the question. It guesses before the model has seen anything, and what it fetches fills the prompt every time.

What Native Context Memory does instead

Nothing is fetched, so nothing is guessed.

Not finetuning

What it does

Finetuning changes the weights and produces a new model. Slow, expensive, and it cannot be taken apart again.

What Native Context Memory does instead

The model stays as it is. The knowledge can be added, edited and removed.

Not memory in an app

What it does

Chat products keep their memory inside the app. Change tool, and you start over.

What Native Context Memory does instead

It lives in the API, so every tool you connect to the same workspace and model has the same memory.

The short answers.

Usually just the base URL and key. The API is OpenAI-compatible, so tool calls and streaming work unchanged. Turn off your harness’s own compaction, so it doesn’t summarise the history the memory keeps. Your agent can keep sending the history as it does today; only what is new is processed.

No. We do not train models on what you send or on what they answer. The model stays as it is, and what you add stays yours to see, edit and remove.

Your work. It goes deeper into the context you have taught it, exactly where you need it.

In European data centres, under GDPR and the EU AI Act. If a model ever runs outside Europe, its listing says so before you choose it. The compliance page has the rest.

Give every tool you use the same memory.

Your agent, your editor and your own scripts draw on one memory. It keeps what you have worked on and what you have taught it, so the next session picks up where the last one left off. You decide what goes in, and you can edit or remove it whenever you like.

One memory, every tool
  1. Pi · Monday · savedLaunch: Friday 16 October
  2. Hermes · Thursday · recalledLaunch: Friday 16 October

Nobody pasted it into Hermes. Both tools read the same memory.