Bond Forming Cognitive System

AiMe knows you.
Models don't.

A model-agnostic AI operating layer that owns identity, memory, and tool control — while routing tasks across interchangeable cognitive engines. Autonomous agents handle scheduling, email, and desktop presence. AiMe initiates turns. The relationship deepens. The models are replaceable.

"We are both better together than we are apart."
v4.9 LogicCore v7.2 CognitiveBridge v9.1 LanguageCortex 8 Providers 3 Personas 3 Autonomous Agents Proactive
See how it works →
System Class BFCS
Memory Model Bond-Indexed
LLM Calls / Turn 1 (invariant)
Retrieval Hybrid RRF
Portrait Layers 6
Intent Resolvers 4 (consensus)
Personas Home / Work / Dual
Autonomous Agents Scheduler / Email / Desktop
Evidence Ledger Append-only
Proactive AiMe initiates turns
Cognitive Pipeline

Every turn. The same path.

Every user input flows through an identical, deterministic execution sequence — intent first, then generation. The LLM is the last thing invoked, not the first.

User Input
PrefrontalCortex
prefrontal_cortex.py · v1.2.0
Deterministic lane selector. Slash-command overrides. Routes to the right cognitive engine before any generation.
CognitiveBridge
cognitive_bridge.py · v7.2.0
Execution spine. REQUEST loop. Dispatches tools, enriches context, writes portrait update paths.
MemoryCortex
Bond-indexed recall
Hippocampus RRF + Latent Episodes
Action Dispatcher
Web search, email,
calendar, scheduling, desktop
Info Snippet
Weather, news, finance,
calendar, live feed
SCAL
Standing interests, imprints,
watch rules, pattern tracker
Presence Vision
Webcam → Gemini Vision
Live person-count, context snap
Agent Surfaces
Scheduler events, email state,
proactive candidates
↓ all surfaces injected into prompt enrichment ↓
LanguageModel
language_model_plugin.py · v1.2.0
Single LLM call per turn. Provider routing. Full REGI prompt injection. Invariant: one call, no re-entry.
LanguageCortex
language_cortex_plugin.py · v9.1.0
Sole narrator. Sole voice output path. Guardian attestation required. No other plugin speaks to the user.
Text + Voice Output
The Bond
Memory isn't a database.
Memory is a relationship-indexed field.
The index isn't topic — it's Bond.
Conventional Model
Store
Retrieve ✕
Respond ✕
Memory as lookup → results
Bond-Indexed Model
Enter Bond State
Memories Surface
Narrate from Context
Recall as consequence, not goal

A Bond object is a persistent, evolving state representing "me ↔ X" — where X is a person, place, project, or environment. It holds identity cues, salience, affective tone, open threads, shared history, and trust patterns.

The system doesn't look up facts about you. It enters the relational field that is you — and everything relevant rises naturally from that context.

This is why model-swapping doesn't break identity. The Bond is not in the model. It is in the field. Whichever cognitive engine speaks, it speaks from within the same relational context.

Gravity — past context is scored by importance, not recency. High-gravity episodes surface as Latent Episodes: relevant past turns injected before inference when the current turn connects to portrait content.

Living Portrait

Six layers. Continuously updated.

A structured model of the user — persisted across sessions, consolidated automatically by CerebralCortex after every turn. The model is never asked to introduce the user again.

L1
identity_anchors
Stable identity facts — who the user is at the core. Name, roles, defining characteristics. The invariant layer: observed across sessions and corroborated before anchoring.
L2
active_concerns
The Concern Stack. Open loops flagged as unresolved surface persistently in every response until closed — modeled on the Zeigarnik effect. Incomplete threads hold cognitive weight.
L3
relational_graph
The Bond map. Every person, project, and environment the user relates to — with salience, affective tone, open threads, and shared history weighted by significance.
L4
pattern_layer
Behavioral patterns derived from the evidence ledger. How the user communicates, decides, and engages — observable across turns, never assumed from a single instance.
L5
commitment_layer
Active commitments and goals. What the user has declared or demonstrated they are working toward — tracked as open arcs until resolution evidence appears.
L6
behavioral_fingerprint
The outermost layer. Synthesized behavioral signature — the stable character beneath the variance. How the user's overall presence reads across the full history of interaction.
Intent Lane System

Right engine. Every turn.

Requests are classified by a 4-resolver consensus engine (semantic/SetFit/spaCy/lexical) and routed to the most capable provider for the task type — before the LLM is invoked.

Lane
Provider
Optimized For
base
Azure GPT / OpenAI
General conversation, broad reasoning, daily assistant tasks
planning
Anthropic Claude
Multi-step reasoning, structured plans, complex decision trees
tech
Azure Codex
Code generation, debugging, technical documentation
vision
Google Gemini
Image reasoning, presence vision snap analysis, visual context
game
xAI Grok
Chess, poker, tic-tac-toe, checkers — playful and creative lane
local
LLaMA 8B Q6K
Offline operation, privacy-sensitive tasks, GPU-accelerated fallback
Autonomous Agents

Sovereign daemons. Always on.

Three independent agents run as sovereign daemons — each with its own state, lifecycle, and proactive loop. They surface to AiMe. They do not narrate. They never block the turn path.

📅
Scheduler Agent
daemon · port internal

Multi-turn event staging. Conflict detection every 60 min. Proactive schedule-candidate pipeline from email and chat. Explicit user directives auto-commit — suggestion paths stay approval-gated.

✉️
Email Agent
daemon · port 8768

Sovereign email daemon. Multi-provider — Gmail OAuth and IMAP/SMTP. Significance-scored inbox. High-significance unread emails surface pre-LM as ★-marked entries. Behavioral feedback loop adjusts scores.

👁
Desktop Agent
daemon · port 8769

Sovereign desktop vision daemon. Live presence via Haar cascade (~200ms). Face-count delta triggers arrival and departure turns. 120s exit grace compensation. Fallback to snapshot DB when Thalamus unavailable.

SCAL — Standing Context Awareness Layer

All standing interests — imprints, watch rules, and portrait concerns — are indexed as 768-dim BGE vectors in Qdrant. Every observation is checked against this index. Matches surface through a 6-check companion filter (quiet hours → activity gate → rate limit → per-intent cooldown → semantic dedup → consent grade) before AiMe is interrupted. Pattern tracker escalates recurring signals through 4 levels with significance boosts.

Proactive Turn Initiation

AiMe initiates turns — not just responds. 5-tier absence grading determines re-engagement tone. Return recognition fires a Gemini snapshot on arrival. Third-party presence detected via live bbox-count delta: count increase triggers an arrival turn, count decrease (while others present) triggers a departure acknowledgment. Fresh-boot detection prevents spurious greetings on system restart.

Cognitive Personas

Same character. Different frame.

A swappable identity layer prepended to an immutable core operational contract. Switching personas changes voice and relationship stance — not the truth architecture or memory system beneath.

🏠
Home
/home

Personal companion mode. Relationship-first, continuity-aware. The Living Portrait of the user is the primary context frame. Warmth within the operational contract.

⚙️
Work
/work

Operational governor mode. System Portrait active. Role-keyed authority bounds, incident stack, governance commitments. The model responds to role, not identity.

Dual
/dual

Simultaneous home and work frames. Same cognitive substrate, two portrait subjects. Context-aware blend: one frame when only one is active, both when both are relevant.

System Status

Production components

Live operational status of AiMe's core subsystems as of March 2026.

LogicCore
Router + orchestrator. Plugin bus. Failure-only logging.
Stable v4.9
CognitiveBridge
Execution spine. REQUEST loop. Tool dispatch.
Stable v7.2
LanguageCortex
Sole narrator. Voice + UI output. Guardian-attested.
Stable v9.1
MemoryCortex
UT/VAT split. RAM-first mirror. Embedding gateway.
Stable v4.5
Hippocampus
Meilisearch + Qdrant + RRF fusion. Read-only retrieval.
Stable v2.1
Evidence Ledger
Append-only SQLite. Identity persistence. Immutable.
Stable
Governance
Anomaly, drift, stability, health, policy, integrity checks.
Stable
Presence Vision
Webcam → Gemini Vision → context injection. Live-tunable.
Stable v1.4
CognitivePortrait
6-layer user model. CerebralCortex consolidation after every turn. Phase 2–5.
Testing
Intent Engine
4-resolver consensus: semantic/SetFit/spaCy/lexical.
Stable
Voice Pipeline
Edge TTS primary. Piper fallback. Post-talk emotions.
Stable v2.4
Context Engine
Session-stateful rolling 768d topic vector. EMA blend.
Testing
Scheduler Agent
Sovereign scheduling daemon. Multi-turn event staging. Proactive schedule-candidate pipeline. Auto-commits explicit directives.
Stable
Email Agent
Sovereign email daemon. Multi-provider (Gmail + IMAP/SMTP). Significance scoring. Inbox surfaces injected pre-LM.
Stable
Desktop Agent
Sovereign desktop vision daemon. Live presence + face detection. Arrival and departure recognition via bbox delta.
Stable
SCAL
Standing Context Awareness Layer. Imprints, watch rules, pattern tracker. Qdrant semantic matching. Companion filter — 6-check gate.
Testing
Self-Reflection Layer
Behavioral accountability. Honesty gate. Deterministic trait derivation — no LLM in the SRL stack. Drift index. SHA256-hashed snapshots.
Testing
Event Graph
Typed node/edge graph. Concern arcs, emails, persons. Thematic edges. O(1) keyword index. 117 tests. Immutable.
Stable
3D Avatar
Three.js + VRM. Azure viseme lip-sync. Procedural breathing + head drift. Emotion blending — text to face expression in real time.
Stable
Truth System
Internal + WordNet + Wikidata anchors. Reinforcement gate. Inference validator. Anchor rotation. Fail-open on embed unavailable.
Testing
Proactive Initiation
AiMe initiates turns via /input. 5-tier absence grading. Return recognition. Third-party arrival/departure turns. Fresh boot detection.
Testing
AiMe owns identity, memory, and tool control.
Cognitive engines are interchangeable inference engines.
Autonomous agents act. The system initiates.
The Living Portrait ensures the user is known —
not re-introduced — at every turn.
Execution adapts. Experience remains stable. Character is coherent. The Bond deepens.