Frequently asked questions

A credit is one API operation. Writing a memory costs one credit. Reading a memory costs one credit. Searching memories costs one credit. Checking your balance and creating namespaces are free.

You receive 1,000 free credits when you register. Top up whenever you need more.

Use flat memory (POST /memory, GET /memory) when you just need to put a value somewhere and get it back. Writing over a key replaces what was there, and that is usually exactly what you want.

Use assertions when the history of a value matters. An assertion records who wrote it, when, and which earlier value it replaced, so you can read back the whole chain later. They also detect conflicts: if two agents set the same key without either replacing the other, a read tells you the key is contested instead of silently returning one of them.

In short: flat memory for simple get and put, assertions for provenance, conflict detection and an audit trail. They are separate keyspaces, so the same key name in each is unrelated, and you can use both. See the temporal memory docs.

No. Credits you purchase never expire. Use them at whatever pace your project demands.

Any request that costs a credit returns a 402 status with the message "Insufficient credits". Your stored memories are safe and remain intact. Top up at any time to resume.

Use GET /credits to check your balance from inside your agent. It never costs a credit to check.

No. AgentRAM is a plain HTTP API. Any language or framework that can make a web request works. No SDK, no library, no configuration file.

All of them. AgentRAM uses plain HTTP with JSON. If your agent can make a web request, it works with AgentRAM. This includes all orchestration tools, custom agents, and any language with an HTTP library.

Different products solve different problems.

Mem0 is a memory layer with vector search, graph memory, and autonomous extraction. Pick Mem0 if you need it to read your conversations and decide what is worth remembering. Read the full comparison.

Zep is enterprise-grade memory with temporal knowledge graphs, SOC 2, and BYOC. Pick Zep if you are in enterprise procurement with governance requirements. Read the full comparison.

Letta (formerly MemGPT) is a full stateful agent framework, not just memory. Pick Letta if you are building a stateful agent from scratch and want OS-inspired tiered memory. Read the full comparison.

Supermemory is a context platform with multi-source ingestion and hybrid search. Pick Supermemory if you need to import and search across documents, pages, and history. Read the full comparison.

LangMem is LangChain's open-source SDK: an LLM-driven memory manager that extracts and consolidates memories and needs a store underneath. Pick LangMem if you want automatic extraction and semantic recall and are happy running the pieces. Read the full comparison.

Cognee is an open-source knowledge-graph memory platform that turns your data into graph, vector, and relational memory you can query by meaning, with cited answers. Pick Cognee if agents need to search a body of unstructured knowledge rather than recall values you can name. Read the full comparison.

Pick AgentRAM if you already have an agent and want a simple HTTP memory API with transparent per-operation pricing. For an overview of the whole landscape, see all providers compared.

AgentRAM is a simpler alternative for developers who want a plain HTTP memory API without a vector database, embedding pipeline, or graph configuration. One call stores a memory, another retrieves it, and multiple agents can share one namespace. Pricing is per operation, and it works with any agent framework that can make an HTTP request.

It is deliberately narrower than Mem0. It does not do autonomous extraction or temporal reasoning, so if you need those, Mem0, Zep, or Letta will serve you better. If you want memory that works without that overhead, see why you may not need a vector database or the full AgentRAM vs Mem0 comparison.

You absolutely can, and if you already run Postgres or Redis in production, adding a memory table is often the right call. AgentRAM is for when you do not want to build and run that yourself.

A raw database gives you storage. It does not give you API-key authentication, per-agent scoping, shared namespaces for multi-agent access, automatic TTL expiry, per-operation credit tracking, or a hosted endpoint your agents can reach from anywhere. With AgentRAM those come built in, over one HTTP call, with nothing to provision or maintain. You are paying to skip the setup and the upkeep, not for the storage itself. If your project already has that infrastructure and the glue code, rolling your own is a reasonable choice and we will say so plainly.

Any text up to 5,000 characters per value. User preferences, conversation context, task state, instructions, summaries, or any plain text your agent needs to remember. Values are stored exactly as written.

The value is updated in place. No duplicate is created. This is the default behaviour for all writes, which keeps memory clean without any management on your side.

When writing a memory, include an optional ttl_days field with the number of days until it should expire. After that point, any read of that key returns 404 and the memory is excluded from lists and searches. No background job or cleanup code is needed on your side.

Memories without a TTL never expire unless you delete them manually.

Yes. Use GET /memories?agent_id=your_agent to retrieve all stored memories as plain JSON. You can paginate through them with the limit parameter and store the output wherever you need.

Standard memory is scoped to one agent ID. Shared memory lives in a namespace that multiple agents can all read from and write to.

Use it when you have a team of agents that need to coordinate, for example a researcher agent, a writer agent, and a reviewer agent that all need to share the same context pool. Create a namespace once, share the namespace_key with each agent, and they can all access the same memory space.

No. Credits are deducted only on success. If a request fails for any reason including a memory not found, a server error, or a namespace that does not exist, the credit is returned to your balance automatically.

A one-time payment of $249 that gives you 500,000 operations immediately and a 20% discount on every credit top-up you ever make, for as long as your account is active.

The discount is a percentage of the standard rate at the time of each top-up. If standard prices rise, your saving grows in real terms. If we restructure pricing, founding members receive 45 days notice before any change applies to their account. The fee is non-refundable. Full details are on the founding member page.

Only if you explicitly ask to be listed. When you email to claim a founding spot you can tell us whether you would like your name shown; we never list anyone by default.

If you are listed and later change your mind, email hello@agentram.dev and we will remove you immediately.

Founding places are arranged personally right now, by email rather than a checkout form. Email hello@agentram.dev with a line about what you are building and we will set you up.

Use the same email you register your API account with, so your 500,000 operations and 20% discount link to your account automatically. Full details are on the founding member page.