Vision · forward-looking
Memory as a navigable space
This page describes where we believe AI memory is heading — a research direction we are pursuing, not a claim about what the product does today. The shipping product is a persistent memory API. Everything below is the longer arc.
The premise
AI agents are powerful — and amnesiac
Modern models are extraordinary pattern-matchers, but every new session starts in an empty room. The knowledge an agent earns in one conversation evaporates by the next. We have built systems that can reason brilliantly and remember nothing.
Today's answers treat knowledge as flat — text to be indexed, vectors to be searched, context to be stuffed into a window. Each is useful, and each hits a wall:
RAG pipelines
Retrieve relevant chunks, but do not synthesize, consolidate, or improve over time. Yesterday's lesson is just another chunk to re-rank.
Long context windows
Scale poorly and get distracted. More tokens is not the same as memory — the model still forgets the moment the window scrolls.
Vector databases
Excellent at similarity search, but a store is not a memory. Nothing forgets, nothing strengthens, nothing learns from outcomes.
The thesis
From flat storage to structured memory
Our working hypothesis is that the path forward is not a faster retriever but a different substrate: memory that is organized, not just stored. Some concepts are more foundational than others; some are recent, some are dense with associations. A flat list loses all of that. A structured memory keeps it.
We believe that representing knowledge with explicit structure — hierarchy, association, and geometric relationships — can support more connected understanding than is practical in flat, text-based systems.
This is a research thesis, stated as one. The mathematical tools exist: hyperbolic geometry naturally captures hierarchy, agent-based modeling describes how knowledge evolves, and MCP gives memory a standard way to plug into any agent. Whether they combine into something that decisively beats flat retrieval is exactly what we are testing — measured, published, and open to being wrong.
What we promise
A grounded vision: promises we can keep
We separate what ships from what is near term from what is open research. The line between them is the most important thing on this page.
Shipping today
A persistent memory API
FastAPI and Postgres with pgvector. Write a memory once, recall it from Claude Code, Cursor, VS Code, ChatGPT, or any HTTP client. Memories associate, consolidate, and improve from feedback. This is the real product — no spatial rendering required.
Near term
Structure over flat retrieval
Richer representations of how memories relate — hierarchy, recency, and association made first-class rather than reconstructed at query time. The graph at graph.mnemoverse.com is an early, honest look at this direction.
Open research
Memory as a navigable space
Whether geometric, spatial representations of knowledge meaningfully outperform flat text for multi-hop reasoning is an open question we are actively researching — not a capability we are selling today.
Principles
What guides the work
Cognitive truth
We aim to model how memory actually works — accumulation, consolidation, forgetting — not how a database thinks it should.
Radical simplicity
Every feature has to earn its complexity. One memory, one API key, one mental model the developer can hold in their head.
Agent-first
Built for AI agents and adapted for humans, not the other way around. The interface is an API and a protocol, not a dashboard.
Open by default
Our research, protocols, and the SLoD paper are public. The direction is meant to be inspected, argued with, and built on.
Open questions
The research questions that matter
We would rather state the unknowns plainly than oversell the destination. These are the questions whose answers will decide whether the spatial direction earns its keep:
- Can structured, geometric representations beat flat retrieval on multi-hop reasoning — and by how much, measured honestly?
- What does an intuitive way for an agent to navigate a knowledge space actually look like?
- What is the right balance between richer representation and the compute it costs?
- How do we keep multi-agent memory safe — monitored, interpretable, and free of covert or deceptive behavior?
The vision is ambitious. The product is real today.
Start with the persistent memory API that ships now. Follow the research as it unfolds — we publish what we measure, including when the results are modest.