kotobase
メニュー

AIエージェントメモリ · GraphRAG

AIエージェントに、つながり続ける永続メモリを。

複雑なグラフ基盤を運用せず、MCPとAPIを通じて、エージェントとアプリが共有するナレッジグラフを構築できます。

ナレッジグラフ · GraphRAG · MCP · マネージドホスティング

無料で開始できます。Hosted Secure Storageにクレジットカードは不要です。

AIに必要なのは、もう一つのチャット履歴ではありません

エージェントはセッションをまたぐと忘れ、検索は分断された断片を返します。Kotobaseはグラフ運用ではなく、アプリ開発に集中できるようにします。

エージェントは忘れる

役立つentity、claim、relationshipをセッションをまたいで保持します。

RAGのchunkは文脈を分断する

似た文章だけでなく、人・出来事・文書・意思決定の関係を表現します。

グラフ運用は複雑になる

まず安全なhosted storageを使い、本番チームで必要になったらmanaged SLAへ移行できます。

なぜKotobaseなのか

永続メモリ

Give agents memory that survives the session instead of isolating knowledge in chat history.

つながった文脈

Store and query entities, claims, and relationships instead of only disconnected chunks.

エージェントとアプリで共有

Expose the same tenant-scoped graph through APIs and MCP.

追跡可能な回答

Preserve graph state and provenance through content-addressed commits.

Vector検索は類似性を見つけ、グラフは関係性を保持します。

従来のRAG is useful for finding related passages. Graph-backed AI adds explicit entities, claims, and relationships so an agent can follow connected context.

従来のRAG

  1. Documents
  2. Chunks
  3. Vector search
  4. LLM

Kotobaseによるグラフ対応AI

  1. Documents · Apps · Data
  2. Knowledge graph
  3. Entities · Claims · Relationships
  4. API · MCP
  5. Apps · AI agents

Kotobase complements retrieval; it does not claim that every workload should replace vector search.

一つのつながったcontext layer

Apps · Documents · Data
Kotobase Knowledge Graph
API · MCP
Apps · AI Agents

tenant単位

Applications and agents access the graph within the same tenant identity boundary.

content-addressed

Signed immutable commit CIDs make the selected graph state explicit and reproducible.

client-contributed

Clients can query, cache, compute, and contribute verified CID-addressed results.

Connect your graph to AI agents through MCP

Give agents a graph tool surface without building one-off retrieval glue for every application.

Your Data
Kotobase Knowledge Graph
API · MCP
Apps · AI Agents

MCP endpoint

https://kotobase.net/mcp

Tools are generated from the same published lexicons used by the application API.

Build GraphRAG on explicit relationships

Model entities

Represent the people, products, organizations, documents, and events your application reasons about.

Preserve claims

Keep assertions and their graph context available across sessions and applications.

Traverse relationships

Use Datalog, SPARQL, or Cypher-style query surfaces for the shape of query you need.

Reproduce context

Bind work to content-addressed graph commits instead of an implicit mutable database state.

Looking for a practical Neo4j alternative?

Kotobase is an option to evaluate when you want a knowledge graph for agents and applications, MCP access, managed hosting, or content-addressed provenance.

Verified Kotobase evaluation factors
Evaluation factor Kotobase
Start Free hosted secure storage; no credit card required
Agent access MCP
Knowledge graph Yes
Managed option Yes — Starter, Professional, Business Critical, Enterprise
Query Datalog, SPARQL, and Cypher-style
Provenance content-addressed commit CIDs
Dedicated option Yes — Enterprise contract

Migration fit, stated precisely

Kotobase is not presented as wire-compatible with Neo4j. Evaluate your data model, query surface, agent integration, and operational requirements before migrating. This page makes no unverified price or performance superiority claim.

Current GraphDB specification and measured limits

Production scope as of 2026-08-15. サポートed syntax, measured evidence, and open gaps are separated so a subset result is not presented as general graph-database performance.

Honest production capability matrix
Area Current production specification
Storage and truth Signed immutable CID commit DAG; R2-backed immutable blocks and conditional graph heads. Durable Objects, D1, PostgreSQL, and an S3 mutable database are not correctness dependencies.
Cypher-style read subset Labels, parameters, property projection, DISTINCT, ORDER BY/SKIP/LIMIT, NOT/AND/OR predicates, one materializing WITH stage followed by MATCH, simple CASE plus CASE WHEN variable IS NULL, and directed/undirected variable-length paths in primary and OPTIONAL MATCH. OPTIONAL traversal runs under the left join and null-fills misses. A 2026-08-15 production probe combined these new operators and returned the expected row through kotobase.net on graph Worker 838a5a03-f307-46f2-99b8-4717a0573096. This remains a bounded subset, not full openCypher or a performance claim.
Planner and joins Path-free basic graph patterns are handed to the peer statistics planner as one BGP: it measures visibility-scoped clause cardinalities or consumes fresh scoped materialized statistics, cost-orders clauses, and passes estimates to join strategy. Hash natural joins and binding push-down are used around OPTIONAL/WITH stages. Variable-length path interleaving still uses bounded heuristics; this is not a general cost model for every Cypher operator.
Cypher writes Authenticated CREATE of standalone nodes with a required non-empty id uses the same graph-policy check, immutable transaction/commit blocks, conditional R2 head CAS, nonce replay protection, and conflict retry path as datom transact. The 2026-08-15 production probe created created/1 with HTTP 200 and read the graph back. Relationship CREATE, CREATE after MATCH, MERGE, SET, REMOVE, and DELETE are rejected.
SQL JOIN read subset Bounded SELECT over datoms and request-scoped external tables. INNER/LEFT/RIGHT/FULL/CROSS and non-equality joins, derived tables to depth 2, GROUP BY, and COUNT/SUM/AVG/MIN/MAX are supported across at most four sources. Datom scans/results stop at 1,000 and intermediate joins at 10,000. External tables are finite scalar rows supplied in the request, not arbitrary network/database connectors. A 2026-08-14 production probe passed FULL external JOIN, grouped derived-table, and LEFT non-equality JOIN shapes; this is functional evidence, not a performance claim. HAVING, CTE/UNION, windows, DML, comments, and multi-statements remain unsupported.
Cold reads CID verification, Cache API/R2 block fetch, query-ordered pack Range support, and bounded client caches. The published 12.5M-datom qualification policy requires three independent empty-client-cache samples and cold network ceilings of 5 s for point lookup and 12 s for two-hop traversal; its receipt measured cold p99 2.231 s and 6.264 s. This is a reproducible engineering envelope for those two shapes, not a contractual SLA or an arbitrary-query guarantee.
Same-graph writes Optimistic conditional-head CAS, full-jitter exponential backoff, 12 attempts by default (operator-bounded 1–16), then HTTP 409 with Retry-After and retry metadata. Writes are not lock-free successes under arbitrary contention; no acknowledged write may be reported before its CAS wins.
Scale evidence Historical LDBC SF-0.1 receipt: 3,013,602 datoms and five of seven Interactive Short queries. IS2's WITH shape and IS7's OPTIONAL path/CASE IS NULL shape now pass end-to-end tests, but the historical latency receipt has not been relabelled or rerun. SF-1 converts to 25,522,216 datoms; no completed 100M-datom production qualification exists.
Comparison claim No general Neo4j, Neptune, GraphDB, or Datomic performance superiority claim. There is no audited LDBC Neo4j result to compare against; published local paired results apply only to their exact bounded workloads.

What the benchmark does not prove

Warm-cache latency does not license a cold-query claim; SF-0.1 does not license SF-1 or 100M scale; five supported LDBC queries do not mean full Cypher. Re-run your own data and queries before migration.

クイックスタート

1. Start free

Create a tenant-scoped hosted graph. No credit card is required for Free.

2. Connect an app or agent

Use the API or point an MCP-capable agent at https://kotobase.net/mcp.

3. Ingest and query

Write entities, claims, and relationships, then query the same tenant-scoped graph from agents and applications.

Developer details

ipfs pin remote service add kotobase https://kotobase.net <JWT>
ipfs pin remote add --service=kotobase --name=my-doc <cid>
ipfs pin remote ls  --service=kotobase

Authenticate writes and tenant operations with a gftd-AUTHN JWT (Bearer) or self-signed CACAO. Public exact-CID reads are live at /ipld/:cid; canonical queries select an explicit signed commit CID rather than a mutable database head. Built on kotoba.

Choose the plan that fits how you operate

Start with Free hosted secure storage. Choose Starter or Professional when you want Gftd Japan K.K. operating the graph. Business Critical and Enterprise are contract rungs for HA and dedicated deployments.

Free

$0

開発者 evaluating Kotobase and small graphs that need a secure hosted place to live.

You carry operational risk

Hosted secure storage for your graph. Encrypted in transit and at rest. No credit card. No availability SLA.

  • Availability: No availability commitment
  • サポート: Public docs and repository issues
  • Included: 512 MiB, 50 pins of hosted secure storage (AuraDB Free–sized graph plus pins)

Operational responsibility details

The operator hosts the graph as encrypted multi-tenant storage and does not ask you to run a peer. Availability is still best-effort: the service is provided "as is" and "as available" with no availability commitment, no service credits, no incident-response obligation, no recovery objective, and no support obligation. That allocation is why this rung is free.

Starter

$25 / month

Individual developers and small engineering teams who need hosted encrypted storage without a $65+ company graph bill.

Gftd Japan K.K. is responsible

Hosted secure storage at the Supabase Pro price. For individuals and engineers, not a company Aura 1 GB bill.

  • Availability: 99.9% monthly SLA (by contract)
  • サポート: Email support · 1 business day · Business hours, JST
  • Included: Hosted secure storage for individual production graphs; included capacity quoted at sale

Operational responsibility details

Gftd Japan K.K. hosts and operates the graph on this plan. The public price is $25/month — the band an individual engineer already pays for Supabase Pro. Checkout is not self-serve until a Stripe Price matching this label exists; do not charge STRIPE_PRICE_STANDARD (¥2,980 / ≈$20) against this rung. The Aura Professional 1 GB company band ($65) is a separate sales quote.

Professional

$132 / month

Production applications that need the Aura Professional 2 GB band, flat, with a 99.9% SLA.

Gftd Japan K.K. is responsible

Aura Professional 2 GB band, flat. Gftd Japan K.K. operates it. 99.9% SLA with service credits.

  • Availability: 99.9% monthly SLA
  • サポート: Email support · 1 business day · Business hours, JST
  • Included: 500 GiB, 10,000 pins included; usage expansion beyond that

Operational responsibility details

Gftd Japan K.K. carries operational responsibility on this plan: it operates, monitors, patches, and recovers the service; it commits to a 99.9% monthly availability SLA backed by service credits; it acknowledges customer-reported incidents within one business day; and it holds itself to the published data-handling and security boundary. Availability is measured by the daily SLO gate, whose receipts are committed to the repository.

Business Critical

$292 / month

Production graphs that need Aura Business Critical's HA and 24x7 band, not only a 99.9% shared endpoint.

Gftd Japan K.K. is responsible under contract

Aura Business Critical 2 GB band as a flat monthly BaaS. 99.95% and 24x7 by contract.

  • Availability: 99.95% monthly SLA (by contract)
  • サポート: Named escalation channel · Quoted at contract · 24x7
  • Included: Contracted HA capacity in the Aura Business Critical 2 GB band

Operational responsibility details

Gftd Japan K.K. quotes this rung against Aura Business Critical's 2 GB floor. 99.95% availability and 24x7 support attach only to a contracted deployment that clears the multi-region write and on-call gates. They are not offered on the shared production endpoint.

Enterprise

Custom contract

Financial, medical, government, and enterprise data teams under audit or regulatory obligation.

Gftd Japan K.K. is responsible under contract

99.999% SLA, SOC 2-class compliance response, dedicated infrastructure. By contract.

  • Availability: 99.999% monthly SLA (by contract)
  • サポート: Named technical account manager, dedicated escalation channel · 15 minutes for P1, 2 hours for P2 · 24x7
  • Included: Contracted capacity on dedicated infrastructure

Operational responsibility details

Gftd Japan K.K. carries operational responsibility on the terms of a signed enterprise contract, which can exceed the published ones: a 99.999% monthly availability SLA with a service-credit schedule, 24x7 named support with a 15-minute P1 acknowledgement, dedicated tenant and infrastructure, counsel-approved privacy terms where available, data-residency commitments, customer-held key custody, read audit, and agreed retention and deletion policy. These commitments attach to the dedicated deployment the contract provisions; they are not offered on the shared production endpoint.

SLA and operational responsibility details

The binding version of these statements is the 利用規約 of Service (§6 plans, §7 service levels, §8 limitation of liability); the engineering detail is in docs/PLANS.md.

Enterprise trust and security

99.999% availability SLA

A 26-second monthly error budget, with service credits: below 99.999% → 10% of the monthly fee; below 99.99% → 25% of the monthly fee; below 99.9% → 50% of the monthly fee; below 99.0% → 100% of the monthly fee, plus termination for cause without penalty.

Available only by signed contract on a dedicated deployment that clears the readiness gates in docs/PLANS.md; it is not offered on the shared endpoint.

プライバシー and compliance response

Available today: Security questionnaire and audit response.

Legal review required: Data Processing Addendum (DPA), Business Associate Agreement (BAA). These are not available for execution until qualified counsel approves them.

In preparation: SOC 2 Type II, ISO/IEC 27001, ISMAP (Japan government cloud registration). No report or certificate has been issued yet.

Dedicated deployment

Data residency (in-jurisdiction cold pins); Read audit (universal read receipts); Customer-held key custody (HYOK/KMS).

サポート

Named technical account manager, dedicated escalation channel · 15 minutes for P1, 2 hours for P2 · 24x7

More ways to use Kotobase

Obsidian → Knowledge Graph

Keep Obsidian as your local-first writing space and Markdown as the source of truth. Sync notes, frontmatter, and [[wikilinks]], then query the vault as a graph.

content-addressed durability

Pinned commit blocks can be archived as CAR files to an S3-compatible immutable object store. Providers improve availability but do not choose graph truth.

Frequently asked questions

Is Kotobase only for AI agents?

No. Applications and agents can query the same tenant-scoped knowledge graph through APIs and MCP.

Does Kotobase replace vector search?

Not necessarily. Vector search finds similar content; a knowledge graph preserves explicit entities, claims, and relationships. They can be used together.

Is Kotobase compatible with Neo4j?

Kotobase offers a Cypher-style query surface, but this page does not claim wire compatibility. Validate your schema and queries before migrating.

Can I start without a credit card?

Yes. Free hosted secure storage has no payment checkout. Professional is the live self-serve subscription. Starter, Business Critical, and Enterprise are quoted by sales.

ドキュメント

エージェントに持続する文脈を。

エージェントとアプリのための、つながったナレッジグラフから始めましょう。