PROOF
THE PROBLEM
The context window is an agent's only built-in memory. When the session closes, so does everything it learnt.
A vector store appends. When a fact changes, both versions score well and nothing says which won, so the model guesses.
Retrieval plus an LLM call on every question charges the same for a repeated question as for a new one, even when no fact behind it has changed.
Agent Memory is not
a vector database
a chat-history blob
a hand-rolled knowledge graph
a context-window manager
an agent runtime
It is one database holding every fact your agent learns, with its source, its corrections and its history.
HOW IT WORKS
Conversations and documents become typed entities, attributes and relations, each row carrying its source, span and confidence.
One reconciler decides what a new fact does to an old one: corrections supersede, and disagreement below the floor is recorded as uncertainty.
A question runs through a four-tier cascade, and a repeated chat question can reuse an answer whose facts have not changed. Where retrieval runs, seven signals fuse into one ranking.
HOW IT BEHAVES
Session 4 · turn 42
Facts written
Tobie works at Acme as CTO
turn:42 · chars 15-38 · confidence 0.91
Acme is an organisation
turn:42 · chars 34-38 · confidence 0.95
Session 9 · turn 43
Facts after reconciliation
Tobie works at Acme as CTO
superseded · valid_until 2026-04-12 · row retained
Tobie works at SurrealDB
turn:43 · chars 12-30 · confidence 0.88
Ask · today
Where does Tobie work?
SurrealDB
cited turn:43
Ask · as of 2026-04-01
Where does Tobie work?
Acme, as CTO
cited turn:42 · superseded 2026-04-12
CAPABILITIES
Conversations and documents become typed facts about entities. Each fact records when it was learnt, its source and its confidence and, where the source says, when it was true. Reconciliation decisions are made in code, so the same input gives the same result.
Turns and documents share one entity and relation graph, so a fact from a PDF and a fact from a conversation link to the same entity, each citing the passage it came from.
Vector similarity, keyword matching, graph relationships and the history of what proved useful are fused into one ranking. A cached chat answer is never reused once a fact it cited is superseded.
Writes, ranked reads and chat answers each record a trace linked to the facts they touched, in your own database, so an answer can be traced back to its sources.
A new value closes the validity of the old one instead of overwriting it, so what the agent believed on any date is one query.
THE NUMBERS
The score is the mean of ten memory abilities across the full 700-question BEAM 1M question set. The headline carries its judge and date, and every run, protocol and vendor comparison is on the benchmarks page.
BEAM 1M score, official judge prompt, event ordering by Kendall tau
gemini-3.5-flash-lite judge · October 2026
Questions scored, the full BEAM 1M question set
35 conversations
Memory abilities scored, with the mean reported above
Each ability is scored separately, and every score is in the chart below
QUICKSTART
Open Studio, create a context and copy its URL and API key. A context is a tenant, and no API reads across contexts.
Install the client, export the connection details and write a fact. Recall returns it with its source.
Add the MCP server to Claude Desktop, Cursor or Claude Code, and the agent gets remember, recall, context, reflect, forget, upload and inspect.
Install the agent-memory client, export the context's URL, key and ID, then write a fact and recall it with its source.
Add the context to the client's MCP configuration, and the agent gets the seven Agent Memory tools.
One command registers the context's MCP server over HTTP.
INTEGRATIONS
Seven MCP tools over streamable HTTP, adapters for agent frameworks and SDKs, all on one REST surface.
HOW IT COMPARES
Ten rows, and the first causes most of the others: one engine, so a fact, its provenance and its history are written in one transaction.
PRICING
Free, no card needed
Choose Sandbox if you:
Key features:
Billed monthly
Choose Lite if you:
Key features:
Billed monthly
Choose Standard if you:
Key features:
Billed monthly
Choose Plus if you:
Key features:
Annual agreement
Choose Enterprise if you:
Key features:
Tokens are counted on every language-model and embedding call a context makes, including background work such as consolidation.
Features
Sandbox
Lite
Standard
Plus
Enterprise
Plans and usage
Free
$30
$300
$1,100
Custom
3 million (one-off)
1 million
10 million
40 million
Agreed with sales
Not available
Overage rates apply
Overage rates apply
Overage rates apply
Agreed with sales
Memory capabilities
Isolation and security
Support
THE PLATFORM
Database
Document, graph, vector, time-series and relational in one engine.
Agent Memory
What an agent learns, with its source and its time, in the same engine.
Cloud
Managed clusters in the regions you choose, scaled on demand.
Studio
Query, explore and design the schema from the browser.
MCP
Every model that speaks MCP reaches the database and the memory directly.
IN PRODUCTION
14,000+
Developers building on SurrealDB Cloud
4M+
Developers building on SurrealDB worldwide
FROM THE TEAMS
SurrealDB gives us a foundation where we can unify semantic search, knowledge graphs, and AI-driven decision making without stitching together multiple systems. Collapsing responsibility into SurrealDB has become our default engineering posture.
VP of Engineering, Later
GET STARTED
