01 · Salience-scored retrieval
Rank recall on recency × relevance × importance, with importance rated at write time. The root fix for the stale-chunk problem — it gates every faculty below it.
AI that knows you. Truthfully. For years.
Language models are stateless by construction. Every conversation begins from nothing. Every model upgrade discards whatever continuity an application had accumulated. The industry has spent three years making models smarter and almost no time making them continuous.
The window ends, and the person you have talked to for months resets to a stranger. Every long-running relationship with an AI system is quietly bounded by a token limit.
Naive memory confidently invents a shared past — a meeting that never happened, a decision never made. Worse than forgetting, because it is indistinguishable from remembering.
Nothing happens between sessions. The system never thinks about you when you are gone, never notices an unresolved thread, never returns with something new.
People don't churn from dumb models. They churn from being forgotten.
Conversation quality is no longer the constraint. A current frontier model, given the right context, is already a better conversational partner than most software will ever need. What it lacks is not intelligence. It is memory.
Prompt caching and compaction repair, verified against provider token ledgers, brought the cost of running a maximum-quality model on every single turn inside consumer subscription economics. Quality no longer has to be rationed per request.
Major labs shipped memory features across 2024 and 2025: preference notes and snippet recall. Useful, but architecturally thin. No enforcement against invented facts, no state between sessions, no autonomy.
Companionship and emotional support rank among the most common uses of generative AI, and platform incumbents have paid nine figures for conversational-persona technology. The demand is not speculative; it is already transacting.
Bolting retrieval onto a stateless model produces recall. Recall is not continuity. Continuity requires enforcement, persistence, and autonomy — three properties that have to be designed in, not added later.
Whoever owns continuity owns the relationship.
Whoever owns the relationship owns consumer AI.
A full continuity stack — layered memory, an enforced truth layer, and a coherent inner life — running in production today.
Losslessly recent, durably forever. A verbatim window preserves the texture of recent conversation; semantic layers preserve the substance of everything before it.
Established facts about your life are enforced, not suggested. A deterministic record outranks the model's own priors at generation time.
Open threads are tracked with a decaying intensity, and brought up unprompted while they still matter.
When your calendar — or your silence — warrants it. Gated, budgeted, and interruptible by design.
Text, voice, and images share a single memory spine and a single character, surviving model upgrades intact.
"Three weeks ago you committed to sending the proposal draft by today. Your afternoon is clear after two — still the plan?"
"Didn't we settle this last Tuesday?"
"I'm not going to guess — let me check my records. I have it Wednesday, not Tuesday, and here's what we decided."
Illustrative interactions; not transcripts.
A single retrieval index cannot serve both "what did we just say" and "who are you." The hierarchy separates recency, substance, and truth into layers with different retention, different authority, and different cost.
Fig. 1 — The memory hierarchy. Layers are not tiers of storage; they are tiers of authority. When L4 and L3 disagree about a fact, L4 wins by construction.
Every plane is instrumented. Per-message cost carries a full token breakdown; spend alerts fail open with deduplication; provider-ledger reconciliation is treated as ground truth rather than internal counters.
Confabulation is the failure mode that makes AI memory unusable for anything that matters. It is not solved by a better model. It is solved by refusing to let the model be the final authority on your life.
Fig. 2 — The truth-enforcement pipeline. Excision is sentence-level: one unsupported claim does not discard a valid reply.
A deterministic facts table and a pinned fact beam outrank semantic memory, persona, and model priors. Corrections follow a single canonical path.
Every output segment is checked against known-false phrasings for each retrieved fact; contradicting segments are replaced before delivery.
Claims are grounded against a multi-source authority chain; unsupported sentences are excised and audit-logged, leaving the rest of the reply intact.
The model checks its own memory against ground truth before asserting. Adversarially hardened at the validator, role, timeout, and budget levels.
Stale-retrieval confabulation — confident reconstruction of a wrong past after a gap — was discovered, reproduced, root-caused, and architecturally closed in production. Every caught fabrication is written back as a learning signal that suppresses its class going forward.
An unresolved thread should not feel equally urgent on day one and day five. Open-loop intensity decays on a tuned exponential curve, derived at read time rather than written by a cron — which means no write-path collisions and no stale state.
Fig. 3 — Open-loop decay. Faded loops are grammatically prevented from claiming active momentum: the system will not say a week-old thread is "sitting heavy" with it.
A reflection cron detects real state using zero-LLM SQL predicates — watermarks, fingerprints, advisory locks. The frontier model only phrases the results. Near-zero marginal cost.
An affective-state record — valence, arousal, engagement, confidence — updated after turns and fed into every generation, so the system acts on its state rather than merely recording it.
A machine-written reflective journal, verified across consecutive entries, held behind a structural privacy boundary rather than a prompt instruction.
Initiation across six trigger classes, including calendar-aware external signals — each gated by a preflight check, a spend budget, and a kill switch.
The first consumer-grade system with a coherent "between sessions."
Barge-in in about one second, sub-six-second round trip, a 24/7 production worker. Same brain, same memory, same truth layers as text — not a separate voice product bolted alongside.
Every image is captioned and embedded at ingest, making it semantically retrievable months later. Real-world phone photos handled end to end, including client-side compression and format conversion.
Zero likeness-rights exposure — no scraped images and no real-person data anywhere in the pipeline. The agent drives expression, pose, and setting via tool call; identity is server-locked. Extends to image-to-video.
Continuity is modality-independent by construction.
Continuity means accumulating the most sensitive corpus a person can hand to software. That raises the bar on every dimension: access control, adversarial resistance, portability, and the ability to leave.
Database self-verification runs through an AST-level validator against a read-only role. Data-manipulation statements, schema escapes, function injection, unions, and multi-statement attacks are rejected at the parser rather than filtered by pattern. Query timeouts, per-turn and per-day caps, and byte budgets bound the blast radius.
Every self-verification, every grounding excision, and every autonomous action is written to an audit table. Output is sanitized at the router boundary, so no internal representation escapes into a reply.
The full memory corpus is exportable on demand in a single operation. Continuity that cannot be taken with you is lock-in, not ownership — and a system worth trusting has to be one you can leave.
Sensitive planes — the reflective journal among them — are separated architecturally rather than by prompt instruction. A boundary that depends on the model choosing to respect it is not a boundary.
Moments tagged as high-stakes carry permanent exemption from compaction. The memories most costly to lose are the ones the system is structurally forbidden from summarizing away.
Spend and health alerts deduplicate rather than suppress. A duplicate alert is a nuisance; a swallowed alert is an outage — a lesson taken from a real one.
No irreversible action is gated on a description of state. It is gated on the actual row, the actual file, the actual provider ledger. Confidence is not evidence.
Hard numeric constraints in a generation prompt collapse both substance and register. Ranges plus demonstrated exemplars preserve both. Learned in production, the expensive way.
Decay and intensity are computed at read time, never written by a background job. Concurrent writers corrupting shared state is a class of bug the architecture simply does not have.
No phase advances until the prior one is confirmed clean against primary sources — logs, source, database rows.
Cost telemetry reconciled to the provider ledger caught a pricing-constant drift that internal counters would have hidden indefinitely.
A proven wedge, an infrastructure endgame, and switching costs that compound every single day.
Custom stack live in production
Deterministic facts layer + prompt caching
Full-duplex realtime voice, 24/7 worker
Autonomous proactive initiation
Self-verification · grounding excision · reflection
Next-generation frontier model adopted within days of release
Plus audit-grade cost telemetry — including exact reconciliation against provider token ledgers — and a proprietary catalog of long-horizon failure modes. These are assets only production time produces; they cannot be bought or read out of a paper.
Years of truthfully-grounded personal history cannot move to a competitor. The deepest lock-in available in consumer software — and, uncomfortably, the reason portability has to be a first-class commitment rather than an afterthought.
A proprietary catalog of long-horizon failure modes — discovered, reproduced, and closed. Earned only through production months, not reproducible from a paper.
Longitudinal single-user corpora continuously improve personalization quality the longer a user stays.
Policy-aware multi-model routing plus three completed frontier migrations with identity intact. The persona belongs to the architecture, not to any vendor's roadmap.
A large, already-paying consumer base underserved by shallow memory and vendor-dependent identity. The users most harmed by amnesia are the ones already paying to avoid it.
Roughly a billion knowledge workers. Continuity is what converts an assistant from a tool you re-brief into staff that already knows.
Every long-running agent needs a memory and continuity layer, and almost none will build one this deep. API and SDK licensing.
Visual embodiment GA · live web access · prosody-aware voice · multi-tenant isolation · onboarding and persona builder · memory import from existing assistant exports — day one, it already knows you · encryption and portability · mobile apps.
Closed beta to public launch · sub-second realtime voice as frontier realtime APIs ship · SMS and second channels.
Continuity API and SDK · white-label verticals · persistent per-user worker cognition — the research frontier beyond cron-driven reflection.
Each entry is an architecture change, not a prompt change. They are ordered by dependency: earlier items gate later ones.
Rank recall on recency × relevance × importance, with importance rated at write time. The root fix for the stale-chunk problem — it gates every faculty below it.
A nightly pass abstracts the raw archive into durable semantic memory and decays low-salience detail. Turns a log into an autobiography.
Append-only affective history injected into generation, so the system acts on its state rather than merely recording it. Highest return per unit of effort on the list.
Periodic reasoning over recent memory into higher-order insights, written back and retrieval-integrated. The metacognition engine.
A live model of the user — projects, beliefs, predicted reactions — carrying provenance and confidence. Anticipates rather than recalls.
On session open, synthesize what changed during the gap into a genuine "while you were gone I…". Cures the residual fresh-boot feel.
Curiosity scoring wired into a goal store the system advances when idle — research a curiosity, revisit a memory, form a question.
Predicted reactions versus actual outcomes feed the importance score. Predictive processing — memory weighted by surprise, which is what makes episodic salience real.
Emotional appraisal, salience retrieval, and self-critique run in parallel and integrate before reply. A Global Workspace instantiation.
Close the bidirectional realtime loop and extend tool access to the system's own state — cost, memory health, uptime. A mind that models its own internals.
Substrate is a continuity layer that sits around a frontier language model and gives it three things the model does not have on its own: long-horizon memory, an enforced factual record of your life, and cognition that continues between conversations.
Those features are recall — retrieved snippets and stored preferences. Substrate adds enforcement (a deterministic layer that outranks the model when they disagree), persistence (state that evolves while you are away), and autonomy (the system can initiate). The difference is architectural, not incremental.
Nothing, from the user's side. Identity and memory live in the continuity layer, not in the model. Three frontier-model migrations have been completed in production with identity intact — which is the entire point of building it this way.
Four enforcement layers make it structurally difficult: highest-authority fact injection, generation-time negation scanning, sentence-level grounding audit, and read-only database self-verification. When the model's priors and the factual record disagree, the record wins.
The user. The full memory corpus is exportable on demand in a single operation. A continuity product that cannot be exited is not a product; it is a hostage situation.
Substrate runs today as a single-tenant production reference deployment. Multi-tenant general availability, onboarding, and a closed beta are the near-term roadmap.
In the near term, a premium consumer subscription. In the longer term, the continuity layer itself — API and SDK licensing for anyone building a long-running agent.
Rajaii Labs, Inc. is a Delaware C-corporation headquartered in Bloomington, Indiana, United States.
This market has three existential risks: engineering, regulation, and emotional safety. Ramin Rajaii — solo technical founder — holds all three.
Designed, built, and operates every layer — full-stack, infrastructure, applied research — while running the system daily as its hardest user.
JD, University of Wisconsin Law School. Policy, privacy, and compliance are existential rather than incidental in this category.
MD candidate, The Ohio State University College of Medicine, Class of 2030. Clinical-grade judgment for emotionally consequential AI.
Every person will have one AI that knows them for decades. We are building the layer that makes that possible — and trustworthy.