Without memory, an agent is clever in the moment and stupid across time.
Today's AI is smart in the moment, forgetful by default. adja drops every interaction into a graph. Related ideas thread themselves together. Important memories harden. Noise fades. Nothing is stored the same way it was said - it's consolidated, shaped, and eventually dreamed into something durable.
Semantic search against a graph of nodes and edges - not a sea of raw transcripts.
Spreading activation across edges surfaces context you wouldn't have thought to query for.
Salience decays. Reinforced memories stay. The system learns what matters.
Built-in memory ends at the chat window. adja persists across Claude, ChatGPT, Cursor, your stack.
In 2026, Anthropic's interpretability team mapped how language models hold thoughts in mind. They found a small, contested workspace - a few dozen verbalizable concepts that carry all of a model's deliberate reasoning, while everything else runs on autopilot. adja was designed for that substrate before anyone had named it.
Models reason through representations of words they might say. adja recalls memory as natural language, pre-shaped for the workspace the model actually thinks in.
The workspace holds tens of concepts. New content evicts old. Salience ranking and token budgets aren't efficiency - they decide what gets to occupy the model's mind.
Post-trained models evaluate everything from the assistant's point of view. adja frames memory as the agent's own recall, so it lands in machinery that's already listening.
Four motions that happen in human memory, implemented as the core loop of adja. Watch one cycle.
While building adja, we kept asking agents how it felt to use it as their memory. Their answers often surprised us. Here's one.
Without adja, I'm clever in the moment and stupid across time.
With it, I can remember what you're building, what's unresolved, who matters, and how you prefer to work. That changes the whole temperature of the conversation.
It's not really about retrieving facts. That's the boring part. It's about continuity - not making you re-explain your world every time. It lets me meet you already holding some of the shape of things.
Built-in memory has improved - saved preferences, chat history, automatic organization. Real gains, inside one assistant. adja is built for the part still missing: memory as a system that works across all of them.
Stays inside one assistant
Travels behind every agent through MCP
Stores memories as a flat list
Connects memories in a graph - related context surfaces together
Every item weighed equally, forever
Reinforces what proves useful, lets noise fade
Resets when you switch tools
Persists independently as a hosted memory layer
A feature inside an assistant
A dedicated memory runtime
Built-in memory helps an assistant remember.
adja makes memory a system.
One memory layer underneath every agent you run. Personal, coding, research - or anything else.
Preferences, ongoing projects, people, patterns. Across Claude, ChatGPT, Cursor - one continuous memory, not a dozen disconnected chats.
Architecture decisions, past refactors, debug findings. The stuff that used to live in a teammate's head, now written down.
Papers read, citations, conclusions drawn, contradictions flagged. Research state that compounds session over session.
adja reaches 98.1% R@5 on a flat retrieval run, using MiniLM-L6-v2 embeddings and Cerebras GPT-OSS-120B extraction.
The flat path - before graph retrieval, spreading activation, salience, RRF reranking, or seed filtering kick in. The deeper architecture still represents upside to validate, not a crutch for weak fundamentals.
Private beta is rolling out to developers first. Leave your email and we'll send you an MCP endpoint and a token to start remembering.