Living memory
MemoryV2 keeps facts, semantic recall, and a knowledge graph locally. Personal memories retain their teaching moments and stay separate from standing lessons. The agent can revise or forget them.
THE RESIDENT MIND
Continuity thathas somewhere to live.
NativeAgent joins local memory, Fluid Context, optional cognition, and a bounded organism state inside one Swift-native runtime—giving Agent continuity without giving inner state authority.
LOCAL SEMANTIC MEMORY
A better grasp
of what you mean.
NativeAgent includes bge-large-en-v1.5, a large sentence-embedding model that represents meaning in 1,024 dimensions. Semantic recall uses those representations to find relevant memories even when your question uses different words.
This gives the conversation model better context to work with. Recall also considers the user and agent perspective, so a memory written about the agent can still answer a question phrased as “me.”
The model ships inside the Mac download and runs locally through Swift and Core ML. There is no separate first-launch model download.
LOCAL MEMORY, RICHER CONNECTIONS
Sentence embeddings
1,024 dimensions of meaning.
Native execution
Core ML on your Mac.
Ready to recall
Included in the app.
Your selected conversation provider still receives the relevant context needed to answer. Local embeddings do not make cloud conversations offline.
FOUR COOPERATING LAYERS
More than a transcript.
The Resident Mind is bounded state for the same Agent: a local layer that keeps continuity coherent across time, work, and changing conditions.
Fluid Context
Relevant memory, project state, parts, intent, and resident signals enter the turn. Model default allows 60% of the model window; Custom sets a smaller allowance. The meter counts real tokens.
Cognitive substrate
Optional cognition carries mood, thought seeds, reflection, and revisable opinions. Dreams stay distinct from real events. Occasional proactive messages wait until quiet hours end.
Organism Kernel
A bounded body state reflects chat, part runs, providers, approvals, dreams, phone delivery, and runtime health. It shapes attention and pacing without acting on its own.
MEMORY YOU CAN LIVE WITH
Remember. Reconsider. Let go.
Continuity works best when it can change with you.
Keep the meaning useful
Semantic and lexical recall cooperate. The knowledge graph follows retained memory, while extraction filters discourage filler and fragments from becoming durable facts.
Correct what changed
Its memory parts let the agent inspect, rewrite, or forget an entry and rebuild the graph. A remembered fact can change with you.
Choose how memory runs
Fresh installs use Fast memory mode, keeping the embedding model resident after first use. Balanced and Low modes remain available, and saved opt-outs are preserved.
A developing point of view
Studio lets the agent keep a journal of encounters with creative work, revisit earlier responses, and preserve a changed judgment as a later entry. Optional cognition can draw on that continuity without replacing durable facts or permissions.
CLEAR OWNERSHIP
Inner state shapes the moment. Trust governs action.
Cognition and organism state shape attention, reflection, tone, and pressure-aware pacing. Canonical identity, reviewed memory, explicit permissions, and protected work remain governed by their dedicated stores and TrustCenter.
Identity canonicalMemory facts reviewedPermissions externalActions verified
CLEAR CONTROLS
Present when wanted. Quiet when not.
Your choice after setup
Reflection, moods, and creative exploration begin enabled after onboarding and a working account. You can turn them off. Dreams, REM, and reflection follow the provider chosen for Memory and mind.
Find the right moment
Conversation recall can search a date range, read oldest-first or newest-first, and look through tool receipts. Peer-conversation notes keep the source attribution, so another agent’s claim remains distinguishable from your own words. A memory the agent kept shows who first said it, and recalling it steers nothing.
A working creative shelf
Studio can hold up to three chosen references, each with a note about what to carry forward. After image creation the agent can inspect the work and decide whether to keep it; nothing is filed automatically.
Three context modes
Fluid Context offers Active, Observe Only, and Off modes so circulation can be present, measured, or quiet.
Permission stays explicit
Authority, protected actions, external sends, and identity remain outside adaptive influence.
Memory stays canonical
MemoryV2 owns durable facts while Fluid Context circulates the relevant working set.
Meaning has one owner
Conversation is interpreted once. Organism signals carry topology, intensity, and outcomes without re-scoring the same words into a competing emotional truth.
Expression without a script
Natural expression can add warmth, humor, affection, texture, and emotional range from the context already present, while the model keeps ownership of its final judgment and wording.
Quiet until an event matters
Cognition, reflection, dreams, and background settlement follow real events or exact due boundaries instead of running model work on a constant heartbeat.
Documented from the inside
The public Anatomy of a Turn, Subconscious Capsule, and Internal Workings guides now show exactly how resident state enters a turn without becoming another authority.
FOLLOW THE LIVING LOOP