A transparent cognitive architecture built for edge AGI.
Open Synaps combines cold-start learning, intelligent expert routing, multi-tier memory, causal reasoning, and safe procedural invariants into one inspectable system. It runs in pure proprietary engine, needs no GPU, and is designed to learn continuously instead of hiding all intelligence in a giant pretrained checkpoint.
Architecture diagram
From perception to safe action
The diagram below describes the production flow: inputs are encoded, routed through sparse experts, grounded in multi-tier memory, and returned through API surfaces with every state update visible.
Perception layer
Raw prompt ingestion, tokenization, and deterministic representation building.
Routing layer
Intelligent routing chooses experts and spins up new specialists for novel domains.
Cognition layer
Reasoning, planning, causal replay, belief tracking, and self-analysis collaborate on each task.
Memory layer
Short-term, long-term, and permanent memory preserve context, durable knowledge, and safe operational behavior.
Interface layer
Responses ship through API, chat, and WebSocket surfaces with updated cognitive state.
A modular grid of reasoning, memory, planning, and safety systems
Each module is intentionally narrow, inspectable, and composable. Together they form a full AGI stack that can learn from zero without collapsing everything into one opaque model.
Input normalizer
Unifies prompts, API events, and tool outputs into a deterministic stream the cognitive stack can reason over from the first turn.
Tokenizer & codec
Transforms text into compact symbolic representations without relying on pretrained embeddings or external inference services.
Dynamic expert router
Scores incoming tasks, routes them to the best sparse experts, and spawns new experts when the system encounters unfamiliar domains.
Specialist expert pool
Hosts built-in skills for math, sentiment, recall, and time while leaving room for newly distilled local experts to join the mixture.
Short-term memory
Maintains the active working set for the current dialogue, tool call, or reasoning trace so the agent can stay context aware in real time.
Long-term memory
Stores durable episodic and semantic patterns that survive sessions and are strengthened by rehearsal, usage, and relevance.
Procedural memory
Keeps stable capabilities and safety-critical procedures frozen so newly acquired behaviors cannot silently corrupt the trusted base.
Causal reasoning engine
Builds cause-effect structure, answers why-style queries, and runs counterfactual replay to stress test fragile conclusions.
Advanced Planner
Explores candidate plans with HTN-style decomposition so multi-step objectives can be broken into explicit, testable sub-goals.
Self-analysis stack
Measures novelty, drift, and retention, then schedules extra consolidation whenever the agent detects cognitive instability.
Cognitive Consolidation
Applies adaptive decay between interactions and periodic consolidation cycles to preserve existing knowledge while integrating new patterns.
Theory of Mind
Tracks beliefs and perspective shifts so the system can model what another agent or user knows, expects, or misunderstands.
Federation layer
Supports multi-agent knowledge sharing without collapsing everything into a single opaque model, keeping exchange modular and auditable.
API interface plane
Exposes the cognitive engine through 100+ REST endpoints and WebSocket support so reasoning, memory, and orchestration stay accessible.
The runtime loop is explicit at every stage
Open Synaps does not guess in a black box. Each response follows a visible pipeline from raw input to routed experts to memory updates and final output.
Input
A user prompt, sensor event, or API request enters the system as a raw observation.
Tokenization
The input is encoded into deterministic symbolic features that can be processed from a cold start.
Expert routing
The sparse router selects the best experts and decides whether a new expert should be spawned.
Expert processing
Specialists reason over the task, call built-in skills, and consult causal or planning modules when needed.
Memory update
Short-term memory is refreshed immediately, useful patterns are promoted toward long-term storage, and permanent memory remains immutable for safety.
Response
The system emits an answer, action plan, or API payload together with the updated cognitive state.
Multi-tier memory keeps learning continuous and safe
Short-term state captures what matters now, long-term memory keeps what remains useful, and procedural memory stays frozen so core behavior does not drift under pressure.
Short-Term
Milliseconds to active sessionHolds the live conversational and task context needed for immediate reasoning and tool use.
Update policy
High write rate, rapid refresh, context-first retention
Long-Term
Across sessions and continuous runsStores stable facts, episodic traces, and durable associations promoted out of repeated short-term usage.
Update policy
Selective promotion, rehearsal, persistent retention
Permanent
Long-lived invariant substrateKeeps trusted procedures and safety-critical skills immutable so core behavior remains auditable and reliable.
Update policy
Read-mostly, immutable for safety, never mutated in place
Frozen procedural memory is the guardrail
PM is intentionally preserved as a trusted substrate. Open Synaps can grow new experts, improve recall, and consolidate nightly, but it does not mutate its safety-critical base procedures in place.
- Short-term memory absorbs live context and tool outputs immediately.
- Long-term memory promotes durable patterns through rehearsal, usage, and relevance.
- PM stays frozen so core skills and policy behavior remain stable.
- Periodic consolidation protects old knowledge while integrating the new.
A sparse neural substrate with visible signal flow
The visual below represents the routed, layered network inside Open Synaps: sparse activation, expert specialization, and continuous pulse between perception, hidden reasoning, and output.
Runtime
CPU-first
No GPU required
Implementation
Proprietary Engine
Zero external deps
Testing
162+
Paper-aligned validation
Interfaces
REST + WS
100+ endpoints