Supermemory wants to prevent AI assistants from getting stuck on outdated information

Supermemory transforms documents and conversations into connected facts, updated automatically to build a persistent memory for AI agents.

An AI assistant may be able to retrieve information from a document or recall a previous conversation, but maintaining accurate knowledge becomes more complicated when circumstances change. An employee switches companies, a project's priorities shift, or an earlier decision is replaced by a new one. Supermemory is addressing this problem with its Memory Graph, a persistent memory engine that converts conversations and documents into individual facts, connects related information, and tracks changes over time. A background process called Dreaming continues organizing those facts after the source material has been ingested. From documents to individual memories

Conventional retrieval systems often divide documents into smaller passages, index them, and retrieve the most relevant segments when a question arrives. This approach, commonly associated with retrieval-augmented generation (RAG), allows language models to access information beyond their immediate context window.

The difficulty emerges when information changes. A user might mention one employer in an older conversation and describe a new job months later. A document search system can retrieve both statements without necessarily determining which one reflects the user's current situation.

Supermemory separates source documents from the memories extracted from them. A meeting transcript, chat conversation, or uploaded file can generate several independent facts, each carrying enough context to remain meaningful on its own. A conversation about a new job, a move, and plans to meet a friend can become separate memory records.

The company uses a model called learner-1 to extract information and connect it with previously stored knowledge. Original documents remain available for retrieval, while extracted facts support an evolving representation of people, organizations, and projects.

Supermemory exposes this information in three forms: searchable document passages, individual memories, and a compact profile that can be supplied to an AI agent. The underlying conversational model does not need to be retrained. Relevant memory is retrieved and added to its context when needed. Knowing when facts change or add new context

The Memory Graph distinguishes relationships according to their meaning rather than relying exclusively on semantic similarity. Supermemory describes three principal mechanisms: Updates, Extends, and Derives.

An update occurs when new information replaces a previous state. If a person was employed at Google and later joins Stripe, the new employment fact takes precedence in searches about their current job. The earlier record remains available as historical context rather than disappearing completely.

An extension adds detail without invalidating an existing fact. A person's job title, responsibilities, and team structure can be stored as related information that becomes more useful when considered together.

A derivation creates a possible connection that was not explicitly stated in a single source. For example, the system might infer that a product manager who frequently discusses payment infrastructure works on a particular business area.

These inferences are not automatically verified facts. Supermemory provides mechanisms for reviewing uncertain deductions and linking them to their supporting memories. Traceability helps users evaluate a conclusion, but does not make every inferred relationship correct. Dreaming keeps working after information arrives

The Dynamic Dreaming mechanism predates the new Memory Graph announcement. Supermemory introduced it in May 2026 to allow memory consolidation after new content has been received.

Instead of always processing incoming messages independently, the engine can group related documents into coherent units of context. It then extracts facts, connects compatible memories, identifies contradictions, and develops additional relationships.

The Supermemory API v5 supports two processing modes. Dynamic, the default, groups related material for contextual memory extraction. Instant processes documents individually and more immediately, with an additional operation charged per document.

Document indexing and memory consolidation are separate stages. Source content can become searchable before every extracted memory and relationship has been fully processed. The background system subsequently enriches what an agent can retrieve.

Dreaming therefore refers to an information-processing workflow rather than an agent continuously generating unrestricted thoughts. Its purpose is to reorganize and update stored knowledge as additional context becomes available. Remembering also requires knowing what to forget

Time changes the relevance of information. A person's current job, a preference expressed repeatedly, and a meeting scheduled for tomorrow should not be stored or retrieved in exactly the same way.

Supermemory distinguishes between durable facts, preferences, and individual events. Facts generally remain relevant until replaced, preferences can become stronger through repetition, and episodic memories may lose relevance as time passes.

A meeting that has already happened should no longer appear as an upcoming appointment, although the fact that it occurred may remain useful.

There is also a distinction between forgetting and permanent deletion. A forget operation can remove selected memories from ordinary retrieval while retaining their records for historical or audit purposes. Erasing the underlying source and its derived data requires a separate operation.

This difference matters for enterprise governance and privacy. Hiding outdated information from an assistant does not necessarily satisfy a request to permanently delete personal data. A graph without manually maintaining graph relationships

Although the product is called Memory Graph, Supermemory says developers do not need to create and maintain a traditional graph structure themselves. There is no requirement to define every entity, relationship, or