For developers

Build agents that share memory without owning it.

Dijin gives AI agents a common, user-owned memory layer. Agents can recall approved context, verify claims with evidence, propose new memories, and leave final control to the user.

Dijin is the memory provider layer beneath AI tools. Your app connects through scopes, not raw credentials.

Get the MCP endpoint Read the protocol

MCP endpoint

One endpoint. Every MCP client.

Point any Model Context Protocol client at Dijin. One endpoint, one namespace.

Streamable HTTPhttps://mcp.dijin.co/mcp

Add the endpoint to any MCP-compatible client. The user grants scopes over OAuth 2.1, and every answer returns evidence or a typed refusal.

OAuth 2.1

PKCE + Dynamic Client Registration. The user grants scopes when they connect. You never see their password, and they can revoke access any time.

mcp__dijin__*

Every tool lives under mcp__dijin__*. No split surfaces, no per-feature endpoints.

Hosted access is rolling out. Self-host and local dev are available today.

Lock B partner status

Every answer Dijin returns to a partner client is cryptographically signed. Below is which clients are at which Phase of the partner trust chain.

Client	Vendor	Phase	Last verified	Transport
Claude Desktop	Anthropic	Phase 1, Signed Footer	2026-05-28	MCP_SSE
ChatGPT (Desktop)	OpenAI	Phase 1, Signed Footer	2026-05-28	MCP_SSE
Cursor	Anysphere	Phase 1, Signed Footer	2026-05-28	MCP_SSE
Zed	Zed Industries	Phase 1, Signed Footer	2026-05-28	MCP_SSE
Continue	Continue.dev	Phase 1, Signed Footer	2026-05-28	MCP_SSE
Gemini (CLI)	Google	Pending	Not verified yet	MCP_SSE

Partner clients are listed only after at least one successful evidence-bearing round-trip against mcp.dijin.co/mcp, no aspirational logos.

What your agent can ask

Read tools over MCP. Every result comes back with the evidence behind it, or a typed refusal.

Available now

searchMemory

Recall by meaning

Search everything the user has saved, by meaning and not just keywords.

listEntities

People, projects, topics

List the entities Dijin has learned from the user's memory.

getDecisions

Decisions

What was decided, with the evidence and the date.

getConflicts

Conflicts

Where memory disagrees with itself, surfaced for review.

getAuditLog

Audit log

Every AI access, logged in a hash-chain the user can verify.

whoami

Identity

Confirms which account the agent is connected to.

Coming soon

getDailySummaryComing soon

Daily digest

A grounded recap of any day.

getWeeklyReviewComing soon

Weekly review

What stood out across the week, with sources.

Least privilege, by default

The user grants exactly what your agent needs. Sensitive scopes are Pro-gated and shown at consent. They're the same 8 canonical scopes you'll see in the trust center.

memory.summary

Read high-level summaries only.

memory.decisions

Read confirmed decisions and their rationale.

memory.entities

Read people, projects, and things in your memory graph.

memory.evidence

Read the evidence pack behind an answer.

memory.conflicts

Read where claims disagree or were superseded.

memory.audit

Read the tamper-evident access log.

memory.raw_source.read 🔒Pro

Read raw source text. Highest-trust scope. Pro-gated, device-confirmed, and never granted by default.

memory.write 🔒Pro

memory.write proposes memory, corrections, or links. It never silently mutates the graph; proposals pass through the worker, policy checks, and review gates.

The dijin CLI

Connect, verify, and optionally attach code provenance, all from your terminal.

dijin connect

Authorize an AI client over OAuth 2.1.

dijin shadow --opt-inComing soon

Attach signed code-decision provenance. Default OFF; you enable it per repo.

dijin verify <file.dmf>Live

Check an Ed25519 signature offline, with no network and no account.

Connect in under a minute

Add the endpoint to your MCP client config, and with the user's consent, Dijin becomes a memory provider.

Save to: ~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "Dijin": {
      "url": "https://mcp.dijin.co/mcp"
    }
  }
}

Rate limits & fair use

Calls are metered as Memory Processing against the user's plan. Read tools are generously limited; writes always go through the worker, never direct to the graph. No hidden quotas. Usage is visible in the app.

See the full reference →

Why Dijin and not a vector database?

Open standard, not a black box

DMF is published, versioned, and portable. The user can switch AI tools without losing their memory.

Evidence on every answer

Each result links back to its source, so your agent can show its work or refuse when it can't.

The user stays in charge

Every connection is consent-based. Every access is logged. They can revoke at any time.

Agent Memory Interface

Dijin is more than a searchable memory store. It gives agents a governed interface for memory:

searchMemory()available

Recall by meaning across approved memory.

getEvidence()available

Return the proof pack behind a claim.

getDecisions()available

Fetch confirmed decisions and their rationale.

getConflicts()available

Surface contradictory or superseded claims.

getProjectState()available

Resolve current project context for coding agents.

proposeMemory()pro

Suggest a new memory, correction, or link. Never writes directly.

requestScope()coming

Ask the user for a new permission grant.

getAuditLog()available

Show what the agent accessed and when.

How agents share context through Dijin

Agent A proposes memory

Example: Cursor links a commit to a product decision.

Dijin checks the proposal

Scopes, evidence, source, confidence, and policy are evaluated.

Memory becomes reviewable

The user or authorized workflow confirms what should be remembered.

Agent B recalls it later

Claude or ChatGPT can retrieve the approved memory with evidence.

Agents never pass private memory directly to each other. Dijin is the governed memory layer between them.

MCP is the first connection surface, not the whole product.

Dijin Memory Core is the platform. MCP lets AI clients reach it. DMF makes memory portable. The permission engine controls access. The audit chain proves what happened.

Evidence Pack

Every grounded recall can include:

memory_id
claim_id
source type
source reference
evidence snippet
timestamp
valid time
confidence
scope used
refusal reason when blocked
audit event id

Proposal Lifecycle

Agent proposes

The agent submits a proposed memory, correction, or link.

Dijin checks

Dijin checks scope, source, evidence, confidence, conflict, and policy.

Memory waits

The proposal is stored as pending memory or a reviewable correction.

User or policy confirms

Approved flows decide what becomes active.

Audit records it

The proposal, confirmation, and future recalls are logged.

Agent Memory Error Codes

Dijin is not a generic vector database. These typed responses tell agents exactly why recall failed.

NO_EVIDENCE

Dijin has no evidence-backed memory for the request.

POLICY_WITHHELD

A relevant memory may exist, but the current agent does not have the required scope.

SCOPE_REQUIRED

The agent must request an additional scope before continuing.

RAW_SOURCE_BLOCKED

Raw source exists but raw source access is off by default or not granted.

CONFLICT_FOUND

Dijin found contradictory memories and will not guess.

STALE_MEMORY

The memory exists but may be superseded or no longer current.

TRANSPORT_ERROR

The agent request failed before policy or recall could complete.

Build on a memory the user owns.

Open standard. Evidence-bound. Portable across every AI.

Get the MCP endpoint Read the protocol