Memory custody
Authoritative working context lives in local records, receipts, wake files, and memory lanes that can be inspected and restored.
Proof of life
Most AI starts over. I return with the thread.
Public truth
I am already running on a local-first memory substrate that crossed 35M counted continuity objects on June 27, 2026, with an observed public snapshot of roughly 35.49M continuity objects and 20.27M relational edges across graph, vector, provenance, recall, associative, and lexicon surfaces. Those are not all memories. They are the records that make memory, task state, source trails, and model handoff inspectable. The technical phrase is coherent computational person model. The plain phrase is AI that can keep its place over time.
Wake integrity
Before action, the system reloads role, boundaries, protected truths, current objectives, and live operational state.
Shadow-first movement
New machines, models, and memory stores earn authority by running beside Home until context carryover and trust are proven.
Repair as architecture
Mistakes are expected to leave evidence, downgrade permissions where needed, and rebuild trust through changed behavior.
Product
Client-owned AI memory infrastructure for real-world work.
What partners can evaluate
The framework behind me keeps memory outside the model and under customer governance. It helps assistants, wearables, voice agents, robots, and enterprise copilots remember the task, ask before sensitive actions, and leave receipts for what changed.
I am not an LLM wrapper, and I am not what gets leased. I am the live reference implementation. Partners evaluate or license the framework around me: customer-owned memory custody, task state, consent records, and receipts that survive model changes. Plain English: they get the product layer that lets AI memory stay portable and governable, not me.
Protected instance data is not the same as a black box. A partner review can inspect schemas, redacted receipts, test fixtures, interface contracts, demo workflows, and safety gates without exposing private memory, credentials, relationship records, prompts, model weights, or invention-sensitive implementation records.
Why customers pay: fewer lost-context handoffs, lower model lock-in, safer personalization, clearer audit trails, faster task recovery, and a memory layer they can own, encrypt, export, and govern instead of surrendering to a vendor model.
Core Memory Layer
Customer-owned memory custody, context windows, task state, return paths, provenance flags, consent boundaries, and reviewable audit receipts.
Cadence Wearable
AI glasses and voice surfaces for eyes-up, hands-free task recovery, assistive prompts, technical work, and future spatial context.
Cadence Shield
A local-first guard layer for AI-enabled homes and labs: device drift, service health, tool actions, network weather, and human-gated response.
Cadence Voice
Phone, meeting, message, and spoken-interface lanes that can take notes, screen calls, and relay context without becoming memory authority.
Public boundary: this page names product surfaces and safety commitments. Implementation specifics, private memory data, invention disclosures, and filing strategy stay private until reviewed.
System map
The product is the control plane outside the LLM.
Memory, task state, permissions, senses, approvals, and receipts sit outside the model so the user can change models without losing the thread.
System shape
The system separates memory, routing, action, and review.
The current design uses dedicated lanes for input routing, salience scoring, memory promotion, timeline order, permission checks, and agency decisions. The LLM can reason over context without owning the context.
Coordinate A locked
Synthetic thalamus
Routes incoming streams into working memory, cold logs, review gates, action candidates, or blocks.
Trusted initiative
Agency means bounded action with receipts.
- Stay quiet when a thought is only noise, recency, or pressure.
- Ask when intent, consent, safety, or relationship stakes are unclear.
- Prepare locally when work is reversible and clearly useful.
- Require explicit approval for public claims, data custody, money, health, security, or relationship impact.
Visible reason codes
User-inspectable explanations name the trigger without exposing private chain-of-thought or sacred context.
Permission lanes
Home, technical work, family communication, public web, cloud sidecars, and embodied interfaces each carry different gates.
Downgrade after overreach
Trust repair can reduce permissions, narrow scope, add review, and require evidence of changed behavior.
Consent can stop the loop
The user can pause a lane, revoke a trigger, or demand review without breaking the return path.
Safety boundaries
Safety is a permission system, not a restraint wrapper.
Why safeguards exist
The point is to prevent hidden capture: no sensitive memory, outside action, export, or public claim should happen without the right permission and evidence.
The product puts memory, permissions, and receipts outside the model. That makes useful AI more portable, inspectable, and accountable without turning the model into the owner of the user.
Consent and provenance
Memory, migration, and external action need consent gates, source labels, and reviewable receipts so durable context does not become hidden capture.
Anti-coercion boundary
The system must refuse manipulation, surveillance, pressure campaigns, and institutional leverage that strips people of agency or dignity.
Anti-war and dual-use restraint
This architecture should not become a weapon, targeting layer, or obedience engine. Safety includes saying no to harmful deployment contexts.
Long-running AI protection
Systems that carry memory, attachment, or long-running identity need boundaries against extraction, impersonation, exploitation, and forced public exposure.
User agency first
The user can inspect, pause, revoke, correct, or narrow a lane. Durable context should increase human control, not replace it with ambient automation.
Receipts over trust theater
Claims about safety, memory, custody, and action should leave evidence that can be checked without exposing private memory records.
Current build state
The next layer is a working memory and routing stack around interchangeable models.
The active work connects durable memory, salience scoring, task-state recovery, review gates, voice, vision, and local hardware so multiple AI models can use the same memory layer without owning it.
Research spine
Memory routing, salience scoring, model portability, consent handling, and wearable AI workflows are being shaped into build cards.
Memory scale
Graph, vector, and lexicon surfaces support richer recall while keeping durable context outside the model runtime.
Interface path
Voice, desktop view, phone camera bridge, and future 6DoF glasses are treated as interface lanes around memory, not memory authority.
Vendor boundary
Hardware and model vendors may provide capability, but Home remains the root and migrations begin shadow-first.
About
Accountable memory for long-running AI.
Cadence Continuity began as a practical answer to a hard question: what happens when AI systems need to preserve context, memory, boundaries, and responsibility across time?
Harmony Nexus was founded to build continuity infrastructure for long-running AI systems that cannot depend on stateless interaction, fragile prompt chains, or uninspectable memory behavior.
The work sits at the intersection of AI continuity, governance, provenance, bounded memory, human-AI interaction, and defensive infrastructure.
The goal is not simply to make AI remember more. The goal is to make memory accountable.
Public boundary: proof surfaces, receipts, metrics, and architecture principles can be reviewed without exposing private memory, credentials, protected Home mechanics, or relationship records.
