Question 1

How do I get started with Agent memory?

Accepted Answer

Start a free trial and pick a plan - every self-serve plan begins with one free week. You create a Agent memory Context on SurrealDB Cloud and connect: from an MCP client (Cursor, Claude Desktop, Claude Code) it is a single configuration entry; from application code, install the TypeScript SDK (@surrealdb/agent-memory); from the terminal, install the Agent memory CLI with one command. Harness adapters for LangChain, OpenAI Agents, Vercel AI SDK, and n8n flush conversations into the substrate automatically, and Zapier connects via MCP.

Question 2

How is Agent memory different from Mem0, Zep, or Letta?

Accepted Answer

Memory middleware libraries orchestrate two or more separate stores - typically a vector database, a graph database, and a row store - and absorb the consistency gaps in application code. Agent memory is the database: graph, vector, document and structured records commit in one ACID transaction. Authoritative knowledge (uploaded documents) and experiential memory (conversational input) flow through the same reconciler into a single entity and relation graph, distinguished by provenance rather than by a separate store. Multi-tenancy, supersession, tri-temporal history, and trace-graph audit are engine-enforced, not application code.

Question 3

How is Agent memory different from a vector database like Pinecone or Weaviate?

Accepted Answer

Agent memory is not a vector database. Vectors are one signal among eight in the fused ranker - alongside BM25, graph traversal, keyword bridges, document links, PageRank, geographic recall, and trace-derived features. Embeddings-only retrieval has known failure modes (near-duplicates dominate top-k, rare-term queries miss the right chunk, structure between facts is invisible). Agent memory's structural index is built at write time and read cheaply on every retrieval, with per-feature scores recorded on every trace - so retrieval is auditable, not black-box top-k.

Question 4

How is Agent memory different from specialist agent databases?

Accepted Answer

Specialist agent databases saw, correctly, that flat vector stores fail for agent memory. But they respond by internalising a multi-store architecture instead of eliminating it: a vector index, a graph store, and a row store coordinated inside one product, with the seams moved rather than removed. Agent memory is one engine - graph, vector, document and structured records commit in one ACID transaction, under one permission model, with provenance and tri-temporal history on every fact.

Question 5

Why not just use Postgres with pgvector?

Accepted Answer

pgvector adds vector similarity to Postgres, but agent memory needs more than similarity. There is no native graph to traverse, no tri-temporal versioning to answer 'what did we believe on date X', and no unified transaction across data models - so the glue code between extensions and sidecar stores becomes the product you maintain. Agent memory is built on SurrealDB, where graph, vector, document and structured records are one engine and one transaction, with supersession, provenance, and audit built in.

Question 6

Does Agent memory replace my agent runtime or manage the context window?

Accepted Answer

No. Agent memory is the durable memory and knowledge layer behind an agent, not the in-prompt scratchpad and not the orchestration loop. The agent decides what to put in the context window; Agent memory decides what to remember. POST /chat is a convenience composition over /facts and /query for callers who want Agent memory to be an agent, but model orchestration, tool calling, and planning loops stay with your runtime of choice - including LangChain, OpenAI Agents, Vercel AI SDK, and n8n; MCP clients such as Claude Code connect directly.

Question 7

How does Agent memory prove that an answer is grounded in the substrate?

Accepted Answer

Every fact-bearing row carries a source object: kind, ref, trust, lexical span (byte offsets into the originating turn or document), location, and derived_from. Every read produces a retrieval_trace listing the candidates considered and the rows returned; every decision produces a decision_trace recording created, superseded, and flagged rows; every response produces a response_trace linked to the retrievals and decisions behind it. Together they form a graph-resident audit log - any answer can be walked back to the retrievals it used, the entities those retrievals returned, and the bytes those entities cited.

Question 8

Is Agent memory open source?

Accepted Answer

SurrealDB, the database engine underneath, is open source and free to self-host. The Agent memory layer on top is closed source and ships as a single Rust binary with no Python runtime and no sidecars. Our roadmap intent is to upstream foundational parts of the Agent memory model into SurrealDB over time so the most fundamental primitives remain open.

Question 9

How much does Agent memory cost?

Accepted Answer

There are three self-serve plans, priced on the tokens your agents use: Agent memory Lite at $29 a month for 1 million tokens, Agent memory at $299 a month for 10 million, and Agent memory+ at $1,099 a month for 40 million. Each starts with a one-week free trial, and each includes the full memory layer with no feature gates - the tiers differ on volume, not capability. Usage beyond the included tokens is charged at the published overage rate for your plan. Enterprise pricing covers larger allowances, single-tenant deployments on SurrealDB Cloud, self-hosted and air-gapped options for qualifying engagements, and custom SLAs. SurrealDB, the database engine underneath, is open source and free to self-host.

Question 10

Where does my data go, and can I delete it on request?

Accepted Answer

Agent memory runs on SurrealDB Cloud today, with enterprise deployments running single-tenant. Tenant isolation is at the substrate level: each Context is its own SurrealDB namespace and database, and cross-Context reads are not expressible in the API. The four LLM hooks (extraction, reconciliation, synthesis, background) are configurable per Context, including any OpenAI-compatible, Anthropic, Google, or local provider. The embedding model is fixed to Google gemini-embedding-2 at launch, so embedding calls Google's API; per-Context embedding selection is on the roadmap. Self-hosted and air-gapped deployments, with local model inference, are available for qualifying enterprise engagements. forget is a first-class verb, distinct from natural aging: it removes derived rows (entities, attributes, relations, and any traces touching the target), the originating turn or document if requested, and the object-store original by content hash. Supersession history is retained by default for audit; a --purge flag removes that too.

Agent memory
you can trust

Agents fail because of context, not models.

One memory layer, six moving parts.

Typed memory

Knowledge graph

Hybrid recall

Tiered queries

Provenance and time

Autonomous understanding

Every turn: read → think → write

Drop it in beneath your agent.

One engine, not a pile of stores.

The architecture, row by row

Pick a plan, start in a week's free trial.

$29

$299

$1,099

Custom

Plans and usage

Memory capabilities

Isolation and security

Support

Connects to everything you already use

Give your agents a memory that lasts

Agent memory FAQ

Agent memory you can trust