Cognitia AI
Your portable memory graph
The system of record for personal AI context — entities, temporal facts, behavioral signals, multimodal identity (faces, voices, images), and cross-source provenance in one unified graph.
You own every byte
Full portability. Deletion controls. Production purge on our standard timeline.
Bi-temporal ledger
Facts supersede on contradiction. Full provenance for every change.
Cross-ecosystem ingestion
Google, Microsoft, Apple, finance, social, and messaging as first-class sources.
Watch a memory graph form and self-correct
This is a fictional persona, Maya Chen, built from email, calendar, social, health, and chat. Press play on the timeline to watch two years of memory assemble itself, supersede stale facts (a job change, a move from San Francisco to Austin), and stay correct. Click any node to see exactly what is known and where it came from — including the photos, voice clips, and AI-generated images tied to each person.
Loading interactive demo...
How the three-tier memory layer works
Connected sources land at three depths. Answers start with the synthesized graph when that is enough, then search the semantic index, then open the original record when you need the exact wording. Across all three, identity is resolved, every claim keeps its source, and newer facts supersede older ones.
Knowledge graph
Synthesized people, organizations, places, relationships, and time-bounded facts. Identity and “who / what / when” questions are answered here first — inspectable, not a chat log.
Semantic index
A meaning-based search layer over connected sources — email, files, chats, calendar, and device captures — so the system can find the right material even when names or dates are incomplete.
Source records
The original artifact behind a hit: the message, file, or capture itself. Used when a summary is not enough, and so you can verify what the graph claims against the source.
Provenance
Every property traces back to the channels that contributed it. You can see what informed a person or a fact, and forget a source without guessing what it touched.
Entity resolution
The same person or organization across email, calendar, chat, and files is merged into one identity — aliases and duplicate records resolved instead of left as separate people.
Supersession
When new evidence contradicts an old fact — a job change, a move, a shifted preference — the graph records the change in time rather than silently overwriting history. You can review conflicts and replay the graph as of any moment.
From connect to answer
Connect sources — email, calendar, files, finance, social, messaging, local device
Extract entities and facts into the graph; index the same material for semantic search; keep the original records
Resolve the same person across channels; attach provenance; supersede stale facts as new evidence arrives
Retrieve graph first, then semantic hits, then the source record when the snippet is not enough
Inspect, contradict, or forget at the fact or source level — or export the entire layer
What memory looks like
A grounded answer is a synthesis plus provenance: entities and facts from the graph, then source cards with snippet, channel, time, and an open link when you granted raw access.
Synthesized answer
Agent ReplayYou are planning a Tokyo trip in April and prefer nonstop flights when the fare stays under about $1,200. You want calm, vegetarian-friendly stays near Shibuya or Asakusa, and you now prefer morning schedules — that replaced an earlier evening-meeting preference.
Graph: Ada Chen · nonstop long-haul preference · morning meetings (superseded evening)
2026-02-10
Re: April Tokyo itinerary
Let’s keep the nonstop even if it is a bit more — under $1,200 still works.
Ada Chen, Kenji Mori
Open original
2026-02-12
Tokyo trip
Apr 4–12 · morning blocks preferred · Shibuya / Asakusa hotel shortlist.
Ada Chen
Metadata only
2026-01-16
Meeting time preference
Evening meetings (valid until 2026-01-15) superseded by morning meetings (from 2026-01-16).
Metadata only
How answers are retrieved
Chat and MCP use the same ladder: synthesize from the knowledge graph when that is enough, search the semantic index when the ask is about a source or needs freshness the graph has not absorbed yet, then open the original record when a snippet is not enough. Grounded replies can show Agent Replay — the entities, facts, and source cards used — without dumping raw history.
Identity and relationships start on the graph
Email, files, chats, and “what just happened” start on the semantic index
Source cards attach after synthesis when you ask for provenance
as_of / since replay the graph as it existed at a moment in time
Use it in any AI — MCP & API
The same memory travels to Claude, Cursor, and other MCP clients under scoped, revocable consent. Clients call Ask (natural-language Q&A) or Retrieve (structured graph payload); they do not pick retrieval sub-tools. You choose per-origin fidelity (off, graph-only, or raw). Writes can queue text to remember; they cannot delete.
unified_memory_agent — Ask: synthesized context; optional evidence for used sources; graph time travel
unified_memory_retrieve — profile + scored entities, facts, and edges; optional evidence
unified_memory_agent_config — per-origin grants and multimodal query availability
unified_memory_store — remember a note (origin mcp); returns a source id
Revoke a client anytime from connected apps
How you interact with your memory
Built-in chat and Memory Explorer help you interact with and inspect your memory. They are not the product — the graph is.