Dijin Memory

By Dijin

Overview

Dijin is the Memory Layer for your digital life and AI agents — an open, owner-signed, cite-or-refuse store of your decisions, action items, entities, and evidence citations. Connect this MCP server to ChatGPT and ask it to recall what you've decided on a project, who moved a discussion forward, what your last release shipped, or any other signal Dijin has already captured (from GitHub, Slack, Teams, your own typed Add Memory entries, and more). ChatGPT calls search to discover records, then fetch for the signed EvidencePack — every answer comes back as Ed25519-signed proof with sources and a /api/verify link. Never a guess. Dijin is read-only on the OpenAI side. The MCP server does not expose memory.write to ChatGPT; all writes happen on Dijin's own surfaces, behind the user's own login.

Tools

fetch

ChatGPT

Retrieve one dijin memory record by id (from search) as a signed, verifiable proof. Returns {id,title,text,url,metadata}: text is the extracted claim plus a user-visible signed footer, url is the proof verify link, and metadata carries the full signed EvidencePack. Call this before composing an answer so every statement is grounded in a verifiable Dijin proof. ChatGPT connector compatibility tool.

getConflicts

ChatGPT

Return groups of claims that contradict each other (same subject + predicate, different assertions). Every participating claim carries its signed EvidencePack — a claim that cannot be cited never appears. Returns empty when no conflicts are detected.

getDecisions

ChatGPT

Return decision-type claims extracted from the user's memory. REQUIRES memory_scope ('personal' | 'team' | 'personal_and_team'). Filter by date range. Backed by kg.claims with predicate matching decision semantics. Note: the scope param here selects voice-vs-code LAYERS (orthogonal); memory_scope selects workspace separation.

getEntityTimeline

ChatGPT

Return one entity's memory over time: periods, each holding the claims recorded in it with their signed EvidencePacks. A claim that cannot be cited never appears, so every count is made of things you can open.

getStarted

ChatGPT

Call this FIRST when the user asks what Dijin is, what it can do, how to get started, or why a memory search returned nothing. Also call it proactively whenever search or searchMemory returns zero results, BEFORE telling the user you don't know — it distinguishes 'this account has no memory saved yet' (a setup state) from 'nothing matched this specific query' (an honest refusal). Never describe Dijin itself as unknown or say 'I don't know' about what Dijin is — call this tool instead and answer from what it returns. Takes no arguments.

getWeeklyReview

ChatGPT

Fetch the pre-computed weekly memory summary covering the 7 days ending on the most recent boundary, or a specified week. Returns content + key_points + source_segment_ids. Empty if no weekly summary has been produced yet.

listActionItems

ChatGPT

List actions extracted from the user's memory, each with its signed EvidencePack. An action that cannot be cited never appears. Ordered by recency.

search

ChatGPT

Search the user's dijin memory for relevant records. Returns {results:[{id,title,url}]} where url is a signed, cryptographically-verifiable proof link. ALWAYS call fetch on a result id to retrieve the verified content BEFORE composing an answer. ChatGPT connector compatibility tool.

searchMemory

ChatGPT

Free-text search across the user's full memory: proof-bound claims (your voice/decision memory, extracted into kg.claims), AI claim subjects/objects, and code-layer metadata (commit summaries, decision titles, test summaries from connected git repos). REQUIRES memory_scope ('personal' | 'team' | 'personal_and_team'). Returns ranked snippets with a source field so the AI knows which layer each hit came from. Use this when the user asks 'where did I see X?' or 'have I talked about X?'. Raw server-side transcript search is retired (no raw transcript text store); this searches the derived, proof-bound graph. Note: the scope param here selects voice|code|claims|all LAYERS (orthogonal); memory_scope selects workspace separation.

whoami

ChatGPT

Return the connected user's identity (user_id + email if available) and the scopes granted to this connector. Useful as a 'is my token valid?' health check and for orientating the AI agent about whose memory it can read.

Links

Documentation Privacy Policy

Similar Apps