Supermemory gives AI agents a memory beyond one conversation

Supermemory supplies a memory layer for AI applications that need to carry useful context between conversations. It extracts facts from source material, connects them over time and retrieves relevant information for an agent’s next task. The current offer centres on its developer platform, plugins and memory integrations. Buyers following an older recommendation should first note that the separate Company Brain and Nova applications were discontinued in September 2026.

01 / ProductMemory is an application service with its own lifecycle

The platform overview presents memory as infrastructure that can work across models and agent environments. That separation matters when a product changes its underlying model: the application can retain its own context store rather than treating the transcript inside one provider as its only history. It does not make every stored fact correct or every model equally good at using it.

The graph memory documentation distinguishes original documents from extracted memories. Documents are inputs such as files or conversations; memories represent facts derived from them. Relationships can update previous information, extend it or derive further facts. The distinction gives developers two different things to inspect when a response is wrong: whether the source was accurate, and whether the extracted interpretation preserved its meaning.

A memory service is most useful when context changes. Consider a user who moves from a trial account to a paid team. Keeping both events is useful for history, but treating the old trial restriction as the current entitlement is a product defect. Supermemory’s temporal approach offers a mechanism for that distinction; an application still needs authoritative account checks before performing a restricted action.

02 / AudienceChoose memory for continuity, not as a substitute for account data

The strongest fit is an assistant where a person otherwise repeats stable preferences or prior decisions. A research assistant might remember the preferred reporting format while reading new documents for each assignment. A support assistant might retain an unresolved question between sessions. These are different retention needs, so they should not automatically share the same namespace, expiry expectations or access permissions.

Compare Mem0 when selecting a dedicated memory layer, and Letta when evaluating how persistent context interacts with the agent runtime itself. Start with the application boundary: are you buying a context service to insert into an existing system, or changing how the agent is built and operated? That choice determines the integration work more clearly than a broad memory benchmark.

A simple structured database may remain preferable for fields with one authoritative value, such as an invoice balance or subscription state. Memory can help explain the history around those fields without becoming their source of truth. Teams should decide which information is eligible for extraction and which must always be fetched from the operational system at the moment of use.

03 / WorkflowProposed workflow: carry research preferences between assignments

Begin with a small demonstration assistant and invented research records. Give each test user a separate namespace, then supply a few explicit preferences and prior research notes. Keep the original inputs in an inspectable fixture so reviewers can compare the stored memories with what was actually stated. Do not start by importing a whole organisational drive; that makes errors harder to isolate.

Add a second conversation that changes one preference and contradicts an earlier assumption. Ask both a current-state question and a historical question. A useful result should distinguish what applies now from what applied earlier, without inventing a reason for the change. Include a statement that is deliberately uncertain, such as a possible future deadline, to see whether uncertainty survives extraction.

The content-type guide describes text and URL ingestion separately from file upload. It also distinguishes standard PDF extraction from the advanced extraction available on Scale and Enterprise. Test a text document and a diagram-heavy PDF separately. A successful text answer does not establish that chart labels or figure relationships were correctly extracted, especially where the advanced path falls back to standard extraction.

Retrieve context before generating the answer, and preserve the returned evidence alongside the proposed response. Ask a reviewer to identify an answer supported by the latest source, an outdated answer and an answer that should have remained unresolved. Measure these cases separately. A single average can hide the particular failure that matters most in a personalised application: confidently remembering something the user never said.

The agent integration guide separates tools for developing against the API from the consumer memory MCP. Choose the integration that serves the pilot rather than assuming every item called MCP exposes the same data. The documentation MCP helps an agent read implementation guidance; that is a different permission decision from allowing an assistant to retrieve a person’s long-term memory.

Finish by exercising correction and removal. Delete or replace a deliberately inaccurate input, then repeat the relevant questions after processing has completed. Record which original material, derived memories and application-side copies remain. The acceptance criterion is an understandable result and a reproducible correction path, not merely a successful ingestion response. This is a proposed evaluation, not a report of tests performed by Sequenced.

04 / PricingMonthly credits cover several separately metered operations

The current pricing page displays dollar-denominated monthly plans and a shared usage rate card. The figures below are the displayed USD offer consulted on 29 October 2026. Included credits are a spending balance, not a promise of unlimited ingestion or an equivalent number of language-model tokens.

The rate card separates memory ingestion, retrieval-oriented content ingestion, queries and additional operations. Unchanged content can avoid repeat token charges when supported updates preserve the document identity, while other operations remain chargeable. Monthly subscription credits reset; purchased top-ups persist. Automatic purchase limits are explicitly not a cap on the total invoice, so budget monitoring must follow total consumption.

For an estimate, replay a representative mix of new documents, small updates, searches and optional processing steps. Count ingestion changes rather than multiplying the entire retained library by every query. Add the agent’s model bill and any local infrastructure separately. This produces a budget tied to the actual product behaviour instead of treating the lowest subscription price as the complete operating cost.

05 / DistinctionsThe deployment choice changes who runs the memory system

The local deployment documentation describes a self-contained server binary and a choice of model endpoints. It also distinguishes the public SDK repository from the separate non-public server codebase. Local deployment therefore should not be described as an entirely open-source copy of the hosted service. Check the applicable licence and support arrangement for the deployment being evaluated. The current pricing page lists self-hosted deployments on Scale and Enterprise; fully air-gapped deployment is an Enterprise option.

Running locally can give a team more control over where data is processed, but it also changes responsibility for availability, backups, model configuration and capacity. A hosted pilot and a local pilot need separate operational acceptance criteria. In particular, an extraction model chosen locally may produce different memory interpretations from the models used by the managed platform.

06 / QuestionsResolve discontinued products and derived-information errors

The September closure notice says Company Brain and Nova stopped on 9 September 2026 while the API, MCP and plugins continued. Their old product descriptions are not evidence of current access. This blueprint covers the active memory engine; organisations seeking a packaged company knowledge application should verify the current interface before planning a rollout.

The consequential quality question is whether a derived memory remains attributable and correctable. Ask how the application handles conflicting sources, stale access permissions and inferred facts that users dispute. Keep those cases in a regression set. Vendor-reported speed or benchmark results may help frame questions, but they do not establish the quality of a particular organisation’s documents or conversational history.

07 / DecisionMake the decision around a specific continuity problem