Agent Memory you can trust | SurrealDB

Memory you can finally trust.

Provenance.

Every fact keeps its history, its permissions and its source, in one database.

Retrieval.

Seven signals fuse into one ranking, and the trace records each one.

One engine.

Memory in the database, not beside it: each fact and its edges in one transaction.

PROOF

Every figure is measured or built into the engine.

THE PROBLEM

Where agent memory fails

Memory ends with the session.The work is done again.

The context window is an agent's only built-in memory. When the session closes, so does everything it learnt.

Retrieval returns both facts.The old one and the new.

A vector store appends. When a fact changes, both versions score well and nothing says which won, so the model guesses.

Every question is asked again.A repeat costs as much as the first.

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

Write, reconcile, read. Reconcile is the hard part. Most memory layers leave reconciliation to application code. Here it is part of the engine.

HOW IT BEHAVES

Corrections supersede. The old fact stays.

Session 4 · turn 42

Tobie was just promoted to CTO at Acme.

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

Actually he moved to SurrealDB.

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superseded · valid_until 2026-04-12 · row retained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 · as of 2026-04-01

Where does Tobie work?

Acme, as CTO

cited turn:42 · superseded 2026-04-12

CAPABILITIES

Five capabilities. One memory you can audit. Each capability states the mechanism behind it, and each has a page that shows how it works.

Typed memory

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.

Knowledge graph

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.

Explore the knowledge graph

Recall

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.

Provenance

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.

Temporal memory

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

BEAM at 1M: 77.2%

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.

77.2%

BEAM 1M score, official judge prompt, event ordering by Kendall tau

gemini-3.5-flash-lite judge · October 2026

700

Questions scored, the full BEAM 1M question set

35 conversations

10

Memory abilities scored, with the mean reported above

Each ability is scored separately, and every score is in the chart below

Overall score, then each ability

Instruction following 93.6%

Preference following 89.5%

Information extraction 88.7%

Contradiction resolution 78.6%

Multi-session reasoning 72.6%

See every benchmark, judge and protocol

Create a context.Its own namespace and database.
Open Studio, create a context and copy its URL and API key. A context is a tenant, and no API reads across contexts.
Remember and recall.Every fact keeps its source.
Install the client, export the connection details and write a fact. Recall returns it with its source.
Connect your agent.Seven tools over MCP.
Add the MCP server to Claude Desktop, Cursor or Claude Code, and the agent gets remember, recall, context, reflect, forget, upload and inspect.

CLI

Install and remember

Install the agent-memory client, export the context's URL, key and ID, then write a fact and recall it with its source.

curl -fsSL https://download.surrealdb.com/agent-memory/install.sh | shexport SPECTRON_URL="<context-url>"
export SPECTRON_API_KEY="<context-api-key>"
export SPECTRON_CONTEXT_ID="<your-context-id>"
agent-memory remember "Sam joined Acme as principal engineer in 2024."
agent-memory recall "what is Sam's role" --json

MCP

Claude Desktop and Cursor

Add the context to the client's MCP configuration, and the agent gets the seven Agent Memory tools.

{
  "mcpServers": {
    "agent-memory": {
      "url": "<context-url>/mcp",
      "headers": {
        "Authorization": "Bearer <context-api-key>",
        "X-Spectron-Context": "<your-context-id>"
      }
    }
  }
}

MCP

Claude Code

One command registers the context's MCP server over HTTP.

claude mcp add --transport http agent-memory "<context-url>/mcp" \
  --header 'Authorization: Bearer <context-api-key>' \
  --header 'X-Spectron-Context: <your-context-id>'

INTEGRATIONS

Runs where your agents run

Seven MCP tools over streamable HTTP, adapters for agent frameworks and SDKs, all on one REST surface.

Claude

OpenAI Codex

Cursor

OpenClaw

Hermes

VS Code

Zed

Grok

Zo Computer

LangChain

OpenAI Agents

CrewAI

Google ADK

Pydantic AI

Strands Agents

EveJS

Mastra

Vercel AI SDK

Cloudflare

TanStack AI

ElevenLabs

LiveKit

n8n

Zapier

Notion

MCP

Agent Memory CLI

JavaScript

Python

Swift

Kotlin

Haskell

Elixir

Dart

REST API

HOW IT COMPARES

Memory in the database

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

Plans for every stage. From a free Sandbox to enterprise.

Sandbox

Free, no card needed

Plan: Free; 3 million tokens included, one-off

Choose Sandbox if you:

Want to try the full memory layer with no commitment

Key features:

Knowledge graph with provenance
Tri-temporal memory
Hybrid retrieval (vector + BM25 + graph)
MCP server, TypeScript SDK and CLI

Plan: $30/mo; 1 million tokens a month included

Choose Lite if you:

Are shipping a first agent with modest usage

Key features:

Knowledge graph with provenance
Tri-temporal memory
Hybrid retrieval (vector + BM25 + graph)
MCP server, TypeScript SDK and CLI

Plan: $300/mo; 10 million tokens a month included

Choose Standard if you:

Run production agents with real usage

Key features:

Knowledge graph with provenance
Tri-temporal memory
Hybrid retrieval (vector + BM25 + graph)
MCP server, TypeScript SDK and CLI

Plan: $1,100/mo; 40 million tokens a month included

Choose Plus if you:

Scale agent fleets or heavier ingestion

Key features:

Knowledge graph with provenance
Tri-temporal memory
Hybrid retrieval (vector + BM25 + graph)
MCP server, TypeScript SDK and CLI

Enterprise

Annual agreement

Plan: Custom; Priced for your deployment

Choose Enterprise if you:

Need your own model, a single tenant or custom SLAs

Key features:

Bring your own model (BYOM)
Single-tenant on SurrealDB Cloud
Self-hosted and air-gapped deployment
Custom SLAs and enterprise support

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

Monthly price

Free

$30

$300

$1,100

Custom

Included tokens per month

3 million (one-off)

1 million

10 million

40 million

Agreed with sales

Usage beyond included tokens

Not available

Overage rates apply

Agreed with sales

Knowledge graph with entity extraction

Included

Conversational memory

Included

Tri-temporal versioning and supersession

Included

Provenance on every fact

Included

Hybrid retrieval (vector + BM25 + graph)

Included

Retrieval, decision, and response traces

Included

Multi-modal document pipeline

Included

MCP server, TypeScript SDK, and CLI

Included

Context isolation (own namespace and database)

Included

Scoped permissions and API-key attenuation

Included

Prompt-injection scanning

Included

Forget verb, and purge for GDPR erasure

Included

Single-tenant deployment

Not included

Included

Bring your own model (BYOM)

Not included

Included

Self-hosted and air-gapped deployment

Not included

Included

Community support

Included

Enterprise support with SLAs

Not included

Included

Your questions,answered

Not yet. SurrealDB, which stores every fact, is source-available under the Business Source License 1.1: free to use, modify and self-host, including in production, and the SDKs are open source under Apache 2.0 or MIT. The memory layer on top is closed source today. The one thing SurrealDB's licence reserves is running SurrealDB as a database service for third parties, where they create and manage their own schemas or tables. Each version becomes Apache 2.0 no later than four years after its release.

THE PLATFORM

Everything an application and its agents know. Five surfaces, one engine.

IN PRODUCTION

Trusted at scale. Samsung, Nvidia, Verizon, Tencent, and Walmart run on SurrealDB.

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.

Justin Foley

VP of Engineering, Later

GET STARTED

Give your agents a memory you can trust. Durable memory, trustworthy recall and a knowledge graph in one system you buy, secure and audit once. Built on SurrealDB.

Memory you can finally trust.

Every figure is measured or built into the engine.

Where agent memory fails

Memory ends with the session.The work is done again.

Retrieval returns both facts.The old one and the new.

Every question is asked again.A repeat costs as much as the first.

Write, reconcile, read. Reconcile is the hard part. Most memory layers leave reconciliation to application code. Here it is part of the engine.

Corrections supersede. The old fact stays.

Five capabilities. One memory you can audit. Each capability states the mechanism behind it, and each has a page that shows how it works.

Typed memory

Knowledge graph

Recall

Provenance

Temporal memory

BEAM at 1M: 77.2%

77.2%

700

10

Create a context.Its own namespace and database.

Remember and recall.Every fact keeps its source.

Connect your agent.Seven tools over MCP.

Install and remember

Claude Desktop and Cursor

Claude Code

Runs where your agents run

Memory in the database

Plans for every stage. From a free Sandbox to enterprise.

Sandbox

Enterprise

Your questions,answered

Everything an application and its agents know. Five surfaces, one engine.

Trusted at scale. Samsung, Nvidia, Verizon, Tencent, and Walmart run on SurrealDB.

Give your agents a memory you can trust. Durable memory, trustworthy recall and a knowledge graph in one system you buy, secure and audit once. Built on SurrealDB.