How Does supermemory's Memory Engine Work with learner-1?

Updated

supermemory is a memory and continual-learning engine for AI agents: raw data (chats, URLs, files, raw data) enters the system, the learner-1 model extracts and "dreams on" that context, stores the result in a vector-graph database, and injects relevant memories back into the agent's context as tokens in real time. The practical payoff is that memory lives outside any single model provider, so it carries across models and harnesses. This explanation is for anyone deciding whether to build on supermemory or evaluating how its memory layer fits an agent stack.

The pipeline, step by step

The flow described on the site has four stages, shown in the site's Figure 1:

  1. Ingest. Chats, URLs, files, and raw data are fed into supermemory.
  2. Learn. learner-1 processes that raw context against the existing memory database, performing four operations: update, merge, infer, and forget.
  3. Store. The resulting learnings go into a vector-graph database — described as dense, interconnected, and growing, with an understanding of time.
  4. Inject. Prior memories are brought back to the model by injecting tokens into its context in real time, via hooks.

The key design choice is step 4: memory is delivered in-context rather than by fine-tuning or retraining the model. That is what makes it model-agnostic — the same memory can be re-injected into any model or harness.

What learner-1 actually does

learner-1 is the component that turns raw context into durable memory. The four operations are worth understanding separately:

Operation What it implies for your agent
Update Existing memories are revised as new information arrives, rather than appended blindly
Merge Related learnings are consolidated into interconnected entries instead of duplicates
Infer New conclusions are derived from context, not just stored verbatim
Forget Stale or superseded memories are dropped, keeping the store from growing unbounded

The "understanding of time" in the database matters here: because memories carry temporal context, update and forget can distinguish what is current from what is outdated. For example, if a user tells an agent their preferred language changed last month, a time-aware store can treat the newer statement as current rather than surfacing both with equal weight.

Why memory sits outside the model

The site's stated rationale is that intelligence is no longer the binding constraint — pre-training and post-training keep pushing model capability, but they don't let a model learn and improve in real time. So learning and memory are separated from model providers and kept interoperable. In practice this means:

  • Switching models or harnesses doesn't reset what the agent knows about a user, task, or tenant.
  • Memory is scoped per user, task, or tenant, so multi-tenant deployments keep contexts separated.
  • You are not locked to one vendor's memory feature.

How you connect to it

supermemory is available through an API, plugins, and MCP. The site also offers a setup prompt you can copy and open in an assistant (ChatGPT, Claude, Cursor, Grok are listed as compatible), which is the fastest path to a first working integration. The site reports it runs in production environments including air-gapped ones — relevant if your deployment can't reach external services.

Scale and plan differences

The site states supermemory already serves 100k+ organizations and over a trillion tokens a month. One concrete feature difference between tiers: Scale and Enterprise include advanced document extraction for PDFs, with OCR and descriptions of figures and diagrams, while Free, Pro, and Max keep standard PDF extraction. If your agents need to read scanned PDFs or diagram-heavy documents, that distinction should drive your tier choice. Current pricing is listed on the site's pricing page; check it there rather than assuming terms.

What to verify before committing

  • Retrieval quality on your data. The pipeline is only as good as what learner-1 infers and what gets injected. Test with your own chats and documents to see whether the right memories surface at the right moment.
  • Forgetting behavior. Confirm that superseded information is actually dropped in your use case, not just merged alongside the old version.
  • Document formats. If PDFs with figures or scans are core to your workload, confirm advanced extraction is included in the tier you plan to use.
  • Deployment constraints. If you operate in a restricted network, confirm the air-gapped option fits your environment before designing around the hosted API.