Memory with meaning.
People do not need more folders. They need a way to preserve what mattered, why it mattered, who it mattered to, and what it should change next. Significance is the part current tools throw away.
Memory for AI and humans
Memory is becoming the interface between humans and agents.
Mnemika is an emerging memory layer where human meaning and machine context meet — saved knowledge, living references, personal significance, and recall that any agent can pick up where you left off.
shared
memory
field
memory
field
Peoplemeaning, stories, identity
Agentscontext, tools, continuity
Knowledgesources, evidence, links
Recallask, connect, remember
Timewhat changes, what remains
Actiondecisions, habits, work
Not another notes app.
Mnemika begins with a simple premise: memory is not storage. Memory is the living layer that lets humans and agents continue a thought across time, tools, and sessions.
Context with continuity.
AI agents are powerful in the moment and fragile across time. Close the window and the context is gone. Mnemika explores memory that survives the session — inspectable enough to trust, structured enough to act on.
A shared recall layer.
The frontier is not human versus AI. It is the handoff: what you know, what your agents know, and how both improve together when they point at the same memory.
Standing on a long tradition.
Mnemika did not arrive from nowhere. The idea of memory as an external, shareable, augmenting layer has been built across eighty years of cognitive science, philosophy of mind, and computing. A short map of the thinkers we read.
1945 · The Memex
A machine to extend memory
The original vision of an associative store of all one's books, records, and communications — mechanized for instant, trail-based recall.
As We May ThinkVannevar Bush · The Atlantic 1949 · PlasticityCells that fire together, wire together
Memory as strengthened connection. The principle behind how associations form — and a metaphor for how a memory graph should grow with use.
Hebbian theoryDonald O. Hebb 1968 · ArchitectureSensory, short-term, long-term
The multi-store model: information moves between stores through attention and rehearsal. A foundation for thinking about what an agent holds, buffers, and persists.
Multi-store modelAtkinson & Shiffrin 1972 · Kinds of memoryEpisodic vs. semantic
The distinction between remembering an event and knowing a fact. Mnemika treats both — the lived trace and the distilled knowledge — as first-class.
Episodic memoryEndel Tulving 1974 · Working memoryThe mind's scratchpad
A central executive coordinating temporary buffers. The closest cognitive analogue to an agent's active context window.
Working memory modelAlan Baddeley & Graham Hitch 1962 · AugmentationAugmenting human intellect
Tools, language, and method as a system for raising what humans can collectively know and do. Mnemika is an augmentation system, not an archive.
Augmenting Human IntellectDouglas Engelbart 1998 · PhilosophyThe extended mind
The claim that mind reaches beyond the skull into notebooks, devices, and now agents. The philosophical license for treating a memory layer as part of cognition.
The Extended MindAndy Clark & David Chalmers 1995 · Cognition in the wildDistributed cognition
Thinking happens across people, instruments, and representations — not just inside a head. A blueprint for human-and-agent teams sharing one memory.
Distributed cognitionEdwin Hutchins 1991 · Situated learningCommunities of practice
Knowledge lives in shared practice, not isolated minds. Memory becomes useful when it circulates and is acted on together.
Situated LearningJean Lave & Étienne WengerThe shape of the product
A memory OS for thought, work, and agents.
We are prototyping workflows where bookmarks, notes, conversations, research, habits, decisions, and agent state become one navigable field — searchable, explainable, and useful when the next action arrives.
Think less “archive.” Think more “continuity engine.”
01
Capture signal
Save the article, transcript, note, or conversation while the context is still alive.
02
Connect meaning
Link source, interpretation, emotion, decision, and future task into the same memory trace.
03
Retrieve with intent
Ask from the human side or the agent side and get back the context that matters now.
04
Act, then remember again
Close the loop: what happened, what changed, what should be surfaced next time.
The ecosystem is already building.
Memory for agents is one of the most active frontiers in open source. These are projects we watch and learn from — the working ground beneath the theory.
Agent memory · MemGPT
Where Mnemika fits
Letta
Stateful agents with long-term memory and self-editing context, from the team behind the MemGPT paper.
github.com/letta-aiLetta (formerly MemGPT) Memory layerMem0
A self-improving memory layer for LLM applications — personalized, persistent recall across sessions and users.
github.com/mem0aiMem0 Knowledge graphsGraphRAG
Graph-based retrieval that structures documents into entities and relationships for reasoning over large corpora.
github.com/microsoftMicrosoft Research Retrieval frameworkLlamaIndex
A data framework for connecting custom sources to LLMs — ingestion, indexing, and query over private knowledge.
github.com/run-llamaLlamaIndex Stateful orchestrationLangGraph
Build durable, stateful agent workflows as graphs — with persistence and checkpointing between steps.
github.com/langchain-aiLangChain Memory engineCognee
Memory for AI agents built on knowledge graphs and vector stores — reliable, structured recall in a few steps.
github.com/topoteretesCognee Personal AIKhoj
A self-hostable second brain — search and chat across your own notes, documents, and the web.
github.com/khoj-aiKhoj Local-first interfaceOpen WebUI
An extensible, self-hosted interface for running models and assistants offline — privacy by default.
github.com/open-webuiOpen WebUIThe connective layer
These tools each hold a piece — graphs, vectors, state, interfaces. Mnemika explores the layer that joins them around human meaning and consent.
— our thesisSwitch perspectives.
Tap between human memory and AI memory. The product lives in the overlap.
Human memory is autobiographical.
It carries emotion, priority, story, body-state, and identity. The same fact has different weight depending on who remembered it and why.
Mnemika should preserve context without flattening it into a database row.
Example memory trace
S
Source
“Article about agent memory and retrieval quality.”
M
Meaning
“This matters for making agents useful beyond a single session.”
A
Action
“Prototype a shared memory map for Mnemika.”
Where it comes from.
Mnemika grows out of work on the brain and on AI memory in Luxembourg — two practices that both ask the same question: how does what we know become what we can act on?
Clinical neuroscience
Neurofeedback Luxembourg
A neurofeedback and qEEG practice studying how the brain learns and self-regulates. Years of work on attention, learning, and brain measurement inform how we think about memory as something that changes over time.
neurofeedback-luxembourg.com ↗Knowledge & recommendation research
Brain Curator
A research project on personalized, evidence-linked recommendations — organizing knowledge so the right context surfaces for the right person. The same instinct behind Mnemika's memory layer.
brain-curator.org ↗Mnemika is an independent, exploratory project. It is not a medical device and makes no diagnostic, treatment, or health claims. Links above point to related work for context only.
The hard parts, named.
A memory layer for people and agents only earns trust if it gets these right. They are the design commitments we hold ourselves to — and the questions worth pushing on.
Privacy by design
Memory is intimate. Local-first and minimal-collection should be the default, not a setting.
Provenance
Every memory should carry where it came from, when, and how sure we are — so it can be trusted or questioned.
The right to forget
Forgetting is a feature of healthy memory. Decay, deletion, and redaction need to be first-class.
Consent-aware sharing
What an agent may read, and on whose behalf, must be explicit and revocable.
Interoperability
Memory should move between tools and models — open formats over a new silo.
Evaluation
How do we measure whether recall actually helped? Memory needs honest benchmarks.
Temporal memory
Beliefs change. The layer must hold what was true then and what is true now without confusion.
Portability & export
You should be able to take your whole memory and leave. No lock-in.
Local-first option
For the most personal memory, the machine on your desk should be enough.
What is an AI memory layer?
As AI agents move from one-off chats to ongoing collaborators, the missing piece is memory — a persistent layer that remembers context, decisions, and meaning across sessions. Mnemika is being built as that memory layer for both AI agents and the humans who work with them.
Human memory and AI
Human memory is autobiographical and meaning-laden; machine memory is operational and structured. The interesting work sits in the overlap: a second brain for AI agents that also serves the person — preserving why something mattered, not just what was said. This is the lineage of the extended mind and distributed cognition, made practical.
Why an agent memory layer matters
Without memory, every agent restarts from zero — re-asking, re-reading, re-deciding. An agent memory layer gives continuity: sources, state, permissions, prior attempts, and the reasoning behind a decision. Done well, it is inspectable enough to trust and structured enough to act on, with privacy, provenance, and the right to forget built in.
A second brain for people and agents
Mnemika treats bookmarks, notes, research, conversations, and agent state as one navigable field — searchable, explainable, and useful the moment the next action arrives. Less archive, more continuity engine. Join the signal list to follow along as it forms.
Opening soon
Join the first circle of memory builders.
For people building second brains, AI agents, research systems, personal knowledge workflows, and tools that help memory become action.
We store only your email to notify you at launch — no tracking cookies, no sharing. Unsubscribe anytime by replying. If the list is briefly unavailable, the form hands you a ready-to-send note instead.