Sentience will be the first and only company in the world to truly mirror the human mind
A frontier LLM with connectors is a head with no hippocampus. We built the rest of the brain.
Every advantage on this page traces back to one architectural decision: we mirror the functional systems of the human mind in software, in the spirit of the Standard Model of the Mind. We are building towards a full emulation of a human mind: hippocampus, basal ganglia, prefrontal executive, and the consolidation cycle that runs while you sleep.
Memory is not a database. Cognition is not RAG.
The dominant agent design today is: large model + tool calls + vector store. Smarter weights, wider context, more connectors. The architecture is otherwise flat.
The human brain is not flat. It is a heterogeneous collection of subsystems — hippocampus, neocortex, basal ganglia, cerebellum, amygdala, prefrontal cortex — each running a different algorithm on a different timescale. Memory alone fractures into at least four dissociable systems. Treating that as one undifferentiated “memory” is an architectural bug, not a simplification.
We took the opposite bet. Build the subsystems. Let frontier LLMs serve as a cortical-level substrate where they fit, and engineer the surrounding architecture — encoding, consolidation, retrieval, executive control — explicitly. The result is a system whose behavior on multi-week, multi-modal, personally-grounded tasks is qualitatively different from anything else on the market, because the architecture is qualitatively different.
The memory architecture
Four dissociable memory systems, each with a corresponding anatomical substrate in the brain and a corresponding implementation in our stack.
Working memory
Dorsolateral PFC · phonological loop · visuospatial sketchpad · central executive
The active scratchpad. In humans, ~7±2 chunks held online by sustained PFC firing — capacity-limited, the bottleneck on most cognitive tasks.
What we built: A 15-minute sliding context buffer holding the last things you saw, said, and did, available to every downstream module.
Episodic memory
Hippocampus · DG → CA3 → CA1 · entorhinal cortex
What happened, when, and where — encoded as discrete spatiotemporally-bound events. The hippocampus does fast pattern separation in the dentate gyrus and pattern completion in CA3; cortex slowly consolidates the trace over days to years.
What we built: Timestamped multi-modal capture in PostgreSQL plus dense vectors in TurboPuffer. Postgres serves the slow-cortex role; the vector index serves the hippocampal cue-completion role.
Semantic memory
Anterior temporal lobes · hub-and-spoke architecture
Decontextualized knowledge — facts, concepts, relationships, decoupled from the moment of acquisition. The ATL is a cross-modal hub; damage produces semantic dementia, the selective loss of meaning while episodes survive.
What we built: An entity knowledge graph layered over LLM parametric knowledge. People, projects, tools, places, relationships — extracted from episodic memory by a consolidation pass, mirroring how semantic memory distills from repeated episodic experience.
Procedural memory
Basal ganglia · cerebellum · motor cortex
Skills, habits, action sequences — implicit, non-declarative, learned through repetition with reinforcement. A Parkinsonian patient loses procedural learning while episodic memory persists.
What we built: Roadmap. The next memory system we encode — learned routines and recurring action sequences compiled into reusable skill primitives that don't require deliberation each time.
The cognitive processes that operate on memory
Storing the trace is the easy part. The advantage comes from how we encode, consolidate, and reconstruct — operations almost no agent system implements.
Elaborative encoding
Levels-of-processing · ventrolateral PFC
Memory strength is set at write time, not at read time.
Depth of processing — semantic elaboration, self-reference, association to existing knowledge — predicts later recall far better than raw rehearsal time.
What we built: Every captured episode passes through an elaborative encoder before storage: it is summarized, tagged with entities, linked to existing semantic-graph nodes, and assigned salience. We don't store raw streams. We store rehearsed, schema-bound traces.
Memory consolidation · sleep-time compute
Hippocampal–neocortical replay · slow-wave sleep · spindles
The brain runs offline jobs while you sleep.
Place-cell sequences from waking experience are replayed at 10–20× speed during slow-wave sleep. SWS replays the day's episodes back into cortex, integrating them into existing schemas, forgetting the noise, and strengthening the signal. Memory is not finalized at encoding — it is finalized offline, on a different timescale.
What we built: A consolidation pipeline that re-reads the day's episodes, re-elaborates them in light of newly-learned context, integrates them into the entity graph, and prunes redundancy. Sleep-time compute: cycles spent improving stored memory while the user is idle.
Cue-driven, constructive recall
Hippocampal pattern completion · default mode network
Recall is reconstruction from partial cues, not playback.
Asked to recall a story across weeks, subjects reshape it according to existing schemas — they don't replay it. The encoding-specificity principle: retrieval works when the cue overlaps the encoded context. CA3's recurrent collaterals do pattern completion — a fragment of input regenerates the full pattern.
What we built: Retrieval that reconstructs answers from many partial sources rather than returning a single matching document. Vector search proposes candidate fragments; a reranker weights relevance, recency, and active-context match (40/30/30); a synthesis pass weaves them into a coherent answer. The output is a reconstruction, not a quote.
Associative & spreading-activation recall
Hebbian potentiation · semantic networks
One memory pulls in its neighbors automatically.
Cells that fire together wire together. Semantic memory behaves like a graph: activating one node spreads activation to associated nodes along weighted edges. Recall is not a point lookup; it is a wave that propagates.
What we built: A query against any one memory automatically pulls in adjacent entity-graph neighbors and co-occurring episodes, weighted by edge strength. This is the difference between answering 'what did we discuss in the meeting?' and answering 'what's relevant to this meeting?' — the second requires spreading activation.
Pattern separation vs. pattern completion
Dentate gyrus (separation) · CA3 recurrents (completion)
Two opposing forces, both necessary.
If every similar event collapses into one trace, you confuse Tuesday's standup with Wednesday's. If no trace ever generalizes, you can't recognize that 'standup' is a recurring pattern. The DG performs pattern separation (orthogonalizing similar inputs); CA3's recurrents perform pattern completion (regenerating wholes from fragments). The tension between them produces both specificity and generalization.
What we built: Sparse high-dimensional embeddings for separation; graph-walked traversals and reranker-driven retrieval for completion. Episodes stay individuated while themes still emerge.
Why this beats the alternatives
Where the architecture diverges from the competition, and why it matters in the limit.
vs. ChatGPT or Claude with connectors (Gmail, Calendar, Drive, MCP)
What they have: A frontier model with the ability to call out to your data sources at request time. Tool use, function calling, retrieval at the edge of the context window.
What they lack: Encoding. Consolidation. A semantic graph distilled from your episodes. Associative spreading activation across modalities. There is no working-memory buffer that persists across sessions; no consolidation pass; no schema-bound traces. A connector is a tool call into a flat data lake. It is not memory.
Why this matters: A connector retrieves what you wrote down. We retrieve what you experienced — including what nobody wrote down. And what we retrieve belongs to you, not to a shared inference engine.
vs. ChatGPT with Memory turned on
What they have: The most widely deployed memory feature in AI: saved facts about you plus reference across past chats, distributed to hundreds of millions of users by default.
What they lack: Scope and ownership. It remembers what you typed into one chat window — not your meetings, messages, screen, or inbox. Its memories are flat notes injected into context: no episodic timeline, no consolidation pass, no entity graph, no working buffer of what you are doing right now. And the memory lives inside OpenAI’s product, on their terms.
Why this matters: Your life does not happen inside a chat box. A model of you has to be built from where your life actually happens — and it has to belong to you.
vs. Limitless / Rewind — capture hardware and meeting memory
What they have: Real capture pedigree: Rewind pioneered continuous screen recording on the Mac, and the Limitless pendant records your conversations with transcription and search on top.
What they lack: Everything after capture. Recording is perception, not memory — search over transcripts is a point lookup into raw traces. No elaborative encoding at write time, no consolidation into a semantic graph, no model of your values, judgment, or voice, and nothing that acts on your behalf.
Why this matters: Capture is the easy first third of the problem. The value is in what a mind does with the trace afterward: distilling who you are from it, and putting that to work.
vs. Personal.ai — a trained personal language model
What they have: The right instinct — one model per person — and a head start on the personal-messaging use case, trained on content you deliberately upload.
What they lack: Depth of signal. A model trained on uploaded snippets learns your phrasing, not your life. There is no continuous multi-source capture, no episodic memory of your days, no consolidation cycle turning experience into structured knowledge about your world.
Why this matters: The moat is not the per-person model; it is the memory architecture feeding it. Voice without memory is an impression of you, not an extension of you.
vs. Microsoft Copilot + Windows Recall
What they have: OS-level distribution: screen snapshots on Windows, Copilot wired into Office, Outlook, and Teams. Nobody has more enterprise surface area.
What they lack: A model of the person. Recall is a searchable screenshot timeline; Copilot is a workplace assistant over tenant data. Neither consolidates experience into a persistent representation of you — and what they do know is scoped to your employer’s tenant, not to you.
Why this matters: An assistant your employer deploys at you is not yours. A digital self has to survive changing jobs, and answer to no one but you.
vs. Autonomous browser/computer-use agents (Operator, OpenClaw, etc.)
What they have: Sophisticated agent frameworks with identity documents, goal tracking, and hundreds of skills. Local execution, open source, highly configurable.
What they lack: A mirror of you. A soul document is a file you write to define an agent — its personality, its values, its rules. It does not learn your values from your decisions, distill your expertise from your work, or internalize your judgment from watching you think. It builds an agent. We build a representation of you.
Why this matters: An agent configured with a soul document is still generic — it knows what you told it, not what you have lived. We capture your experiences, consolidate your knowledge, and reflect your values, expertise, and judgment back to you. The difference is whether you are prompting an assistant or extending yourself.
vs. Connectomics-first whole-brain emulation
What they have: A principled bet that if you scan the wiring at sufficient fidelity, simulation will follow. Bottom-up, methodical, increasingly tractable.
What they lack: The right target. A perfect connectome gives you a high-fidelity simulator of a brain at a single moment in time. But a human being is not a snapshot of neural connections — a human being is experiences, accumulated wisdom, and evolution through time. The map is not the territory.
Why this matters: We took the systems-level bet instead: emulate the functional architecture that makes a mind a mind — the processes that encode experience, consolidate memory, and evolve over a lifetime. That captures what matters about being human, not just the wiring diagram at t=0.