3
Sleep Cycle Phases
4
Memory Tiers
R → E → C
Research · Execute · Capture
3-signal
Hybrid Search
An agent that gets smarter while it sleeps. Cache-aware context with infinite memory and zero context loss. Research-gated execution that thinks before it acts and captures what it learns. State separated from execution. Deploy one Core, scale Engines to match load.
Language models are commoditizing. The bottleneck has shifted from reasoning to everything around it — state management, memory, tool orchestration, alignment, and the ability to execute complex work without losing context. Most agent failures are context failures, not reasoning failures.
Every conversation starts from zero. No persistent memory. No understanding of past decisions. No continuity across sessions.
State and execution fused together. If the LLM crashes, everything is lost. Can't scale. Can't recover. Can't run tasks in parallel.
Toolkits that ship primitives, not systems. No alignment model. No memory architecture. No gating. Assembly required — and most assemblies fail.
Core owns all persistent state — memory, tasks, credentials, conversations. Engines are stateless LLM executors that connect via RPC. Services are thin channel adapters. Each component scales, crashes, and restarts independently.
Most AI systems use guardrails — lists of things the model can't do. Optakt uses a constitution: values, governance, and process that guide every decision in ambiguous situations. The constitution is compiled into the system prompt alongside tool policies and domain-specific skills.
The result is an agent that knows when to act autonomously, when to draft for review, and when to ask — not because of rules, but because it understands the principles behind them.
Like human sleep, the agent cycles through three phases of background maintenance during idle time. Conversations are mined for unrealized insights. The archive is cross-referenced for contradictions. Knowledge base are verified against primary sources. The system gets more coherent every day without operator effort.
Dreaming only. Mines recent conversations for insights, decisions, and reasoning that wasn't captured during live work.
Phases 1 and 2. After dreaming, reflects on the archive — finds contradictions, promotes knowledge to memory, amends stale entries.
All three phases. Overnight, the agent consolidates memory — merges overlapping blocks, verifies claims against live systems, prunes stale knowledge.
Every task passes through two programmatic gates. Before execution, the agent automatically searches six knowledge sources — archive, memory, history, web, codebase, and documents. After execution, decisions and outcomes are committed to long-term memory and searchable archive. Nothing is learned and then forgotten.
Knowledge is organized into four layers, from task-specific to universal. As information proves valuable across conversations, it automatically migrates upward — becoming more persistent, more available, and more efficiently cached. Three-signal hybrid search (full-text, semantic, and knowledge graph) makes everything instantly retrievable regardless of where it lives.
LLM context windows are expensive — every token is billed on every request. Optakt structures its context to optimally map onto each provider’s caching mechanism. Stable knowledge reads from cache at a fraction of the cost. Only changed segments are rewritten. The result: infinite conversation horizon at bounded, predictable cost.
Context grows as conversations progress. Four distinct mechanisms reduce it back, each operating at a different frequency and depth.
Deterministic cleanup of intermediate output. Dramatic savings. No LLM call needed.
Summarizes older conversation into dense anchors while preserving knowledge verbatim.
Deep compression of historical anchors. Promotes proven knowledge to more persistent layers.
Consolidates accumulated changes and cleans up stale references. Runs alongside other events.
The context is structured into ordered segments based on how frequently each type of content changes. Optakt maps these segments onto each LLM provider’s specific caching mechanism — ensuring that stable knowledge is read from cache at a fraction of the input cost, while only actively changing segments are rewritten.
The more stable a segment, the more efficiently it caches. In practice, 60–80% of every request reads from cache — dramatically reducing the per-request cost of maintaining rich, persistent context.
Like human sleep, the agent cycles through dreaming, reflection, and consolidation during idle time. It mines conversations for insights, resolves contradictions in its records, and verifies knowledge against reality. It gets smarter while it sleeps.
Context is structured to optimally map onto each LLM provider's caching mechanism. Stable knowledge reads from cache at a fraction of the cost. Only changed segments are rewritten.
Before acting, the agent automatically searches the relevant knowledge sources for each task. After acting, decisions and outcomes are committed to permanent storage. Nothing learned is forgotten.
Knowledge is organized from task-specific to universal. As information proves valuable, it migrates upward automatically — becoming more persistent, more available, and more efficiently cached.
Three-signal retrieval combining full-text keyword search, semantic similarity, and knowledge graph expansion. Results merged by relevance across all three signals. Every piece of knowledge is instantly retrievable.
Values, governance, and process compiled into the system prompt. The agent knows when to act autonomously, when to draft for review, and when to ask first.
Core owns state. Engines execute. Services connect channels. Each component scales, crashes, and restarts independently. Cap'n Proto RPC between all components.
Scope enforcement, rate limiting, approval queues, schema validation. Deterministic Go code — no LLM cost, no circumvention. Phase-specific tool grants with least privilege.
Secrets are encrypted at rest and never exposed to the LLM. The operator decrypts on startup. Credentials are injected directly into tool execution environments by name — the agent never sees the values.
We deploy tailored AI agents for service businesses. Your workflows. Your data. Your agent.