Context Engineering
Cut prompt tokens by 69% (109KB to 34KB per turn) using schema virtualization, dynamic manifest projection, and state compression.
← All projects
Personal OS Agent Runtime · Solo Developer · Apr 2026 – Pres.
Aiko — Personal OS Agent
Aiko is the agent layer I am building over my own machine: a local daemon, mobile app, Telegram bridge, embodied desktop stage, tool runtime, and memory system wired into one loop. It gives me one interface for files, shell, Git, Docker, calendar, web, voice, reading, image generation, and long-running build work.
Under the hood The brain can move between local Ollama, MiniMax through OpenRouter, Kimi through Moonshot, and a CLI-agent bench. A single persisted identity sits above the model, while 88 manifest-backed skills provide the hands.
System shape Daemon, memory, tool runtime, phone surface, Telegram bridge, and factory orchestration are one loop. The mothership layer is the Skill Graph below: the part that turns scattered builds into callable capability, regardless of which model is active.
At a glance · June 2026
- Callable Skills (DAG)
- 88
- Memory Tiers
- 5
- Limbic Dimensions
- 6
- Brain Routes
- 4
- Retargeted Motion Clips
- 79
October 2026
What changed since June.
5-Tier Memory
Scaled to 11,800+ facts combining episodic hindsight, SQLite FTS5 keyword indexing, and dense embeddings fused via Reciprocal Rank Fusion (RRF).
Autonomous Harness
142 discrete skills executed through deterministic state machines; universal gateway orchestrating Claude Code, OpenAI Codex, Gemini CLI, and local models with automatic failover.
Local Multi-Modal Interface
GPU-accelerated local voice (Qwen3-TTS, Whisper), 3D VRM embodiment in three.js, and offline manga indexing across 245,000+ pages.
The Composition
Four bodies. One mothership.
Aiko is not a single binary. A TypeScript daemon holds the identity; a React Native phone app, Telegram bridge, and desktop stage expose different ways to reach the same system. The body changes. The conversation and tools do not.
Core
aiko-core daemon
The brain. A Linux background process that can hot-swap Ollama, OpenRouter / MiniMax, Moonshot / Kimi, and CLI-agent providers without changing Aiko's identity. It owns memory, tools, proactive work, and event delivery.
- TypeScript
- Node
- SQLite WAL
- Moonshot
- OpenRouter
Mobile
aiko-app
Android (Expo / React Native) client. Reaches the daemon's 100.x.y.z:8080 Tailscale IP and opens
a long-lived WebSocket — zero public internet, zero cloud, encrypted end-to-end by WireGuard.
- React Native
- Expo
- Tailscale
- WebSocket
- MMKV
Bridge
Telegram bridge
A persistent Telegram bot for casual contact and on-the-go control. Destructive actions are gated behind a TOTP challenge; the same shared logger streams ACK events back into the unified pipeline.
- Telegram Bot API
- TOTP
- Auth Gate
- Persistent Session
Presence
Desktop stage
A movable VRM stage driven by Aiko's own stage directions. It connects live chat, Qwen3-TTS voice sync, swappable skins, and 79 retargeted motion clips into a visible PC presence.
- Three.js
- VRM
- Qwen3-TTS
- Kimodo Motion
- X11
Every Turn
The conscious pipeline.
Every inbound message — Telegram or mobile — runs through the same six-stage pipeline. Two of those stages are fast pattern matchers that never touch the LLM. Two background ticks run in parallel: a 60-second decay loop and a 3 AM distiller that compresses yesterday's diary into long-term facts.
01 / Fast
Appraisal
Pattern matcher reads the user message and emits a limbic delta — no LLM call.
02 / Stateful
Limbic Update
Delta applied to persisted state. Mood overlay derived for the system prompt.
03 / LLM
Brain Turn
Dynamic 8-section system prompt assembled per turn; reply streamed back.
04 / Agentic
Tool Loop
Structured tool calls are validated against skill manifests, executed, and fed back into the same turn.
05 / Cost
Response Appraisal
Energy spent. Annoyance updated. Trust adjusted.
06 / Persist
Episodic Write
Turn appended to today's diary. Event mirrored to mobile + dashboard.
60s
Background tick — limbic decay applied, queued proactive messages delivered to whichever surface is connected.
3 AM
Nightly distillation — yesterday's diary read, compressed into a summary, relationship facts extracted and persisted to facts.json.
Memory Stack
Five tiers, five lifetimes.
Memory in Aiko is not one bucket — it is a layered store, each tier with its own lifetime, write cadence, and access pattern. Identity is permanent and read every session; relationship facts are extracted nightly; episodic diaries grow append-only; working memory is per-session; world-state is live.
- 01
Identity
Lifetime · Permanent
Who Aiko is, who the user is, the relationship contract.
persona.md, style doctrine, hard rules. Loaded every session. - 02
Relationship
Lifetime · Long-Term
Preferences, habits, opinions, triggers. JSON key-value store extracted nightly by the distiller from raw diary text.
- 03
Episodic
Lifetime · Append-Only
Daily diary files at
episodic/YYYY/MM/DD.md. Every turn appended. Summarized and compressed at 3 AM the following night. - 04
Working
Lifetime · Session
Last 20 turns. What was just said. Cleared when the daemon restarts; reseeded from the diary on warm boot.
- 05
World State
Lifetime · Live
Machine state, time, environment, connected surfaces, currently-running skill jobs. JSON, rewritten on every tick.
Emotional Substrate
The limbic layer.
Six dimensions. Each one persisted across restarts, mutated by the appraisal engine on every message, decayed on a 60-second tick, and overlaid onto the system prompt as a mood directive. Baseline-locking means that if affection and trust stay high for enough consecutive turns, the resting floor permanently rises.
- Affection
- −100 … +100
- Decay starts 48 h · 1 pt/hr
- Trust
- −100 … +100
- Decay starts 72 h · 0.5 pt/hr
- Annoyance
- −100 … +100
- Decay starts 30 m · 10 pt/hr
- Protectiveness
- −100 … +100
- Decay starts 24 h · 2 pt/hr
- Curiosity
- −100 … +100
- Decay starts 15 m · 15 pt/hr
- Energy
- 0 … 100
- Recovers from 2 h · 5 pt/hr
Proactive engine. Aiko can queue messages of its own. Four tiers — silent (A), queued (B), immediate push (C), and check-in (D) — are gated by current annoyance and focus mode. The 60-second tick decides which queued messages fire, and which surface they fire on. Late-night softness, post-silence check-ins, and focus-mode lockouts are tactical shifts derived from this layer alone.
Skill Graph
88 hands. One system.
The active registry has grown from 54 to 88 manifest-backed skills. Foundations such as file, shell, and process access compose upward into library work, avatar control, diagnostics, media pipelines, and multi-agent pilots. The nodes below are a selected map of the current registry, not a frozen catalogue.
Mothership Definition
Where abandoned projects become working limbs.
This is the mothership: not a folder of old repos, but a core that can wrap them, validate them, route them, and let Aiko use them. A hackathon bot, a CLI utility, a media script, a model pipeline, or a weekend automation stops being a dead project once it ships as a manifest-backed skill. The graph records its dependencies, inputs, outputs, permissions, safety rules, and tier placement, so Aiko can compose it with everything else instead of treating it as another isolated codebase.
- repo → manifest
- script → skill
- skill → graph node
- graph node → callable by Aiko
- skills → composable workflows
FoundationsPrimitives
- shell-exec
- file-read
- file-write
- file-find
- file-move
- file-delete
- dir-list
- process-list
- network-status-get
- system-info-get
- power-set
- service-set
- log-read
- totp-verify
SystemControl & State
- git-run
- docker-run
- docker-manage
- deps-install
- dependency-set
- package-set
- health-check
- job-schedule
- task-schedule
- task-control
- calendar-set
- sandbox-run
Mind + MediaRecall & Creation
- memory-save
- memory-search
- memory-decay
- diary-write
- knowledge-fetch
- text-extract
- text-summarize
- generate-image
- upscale-image
- passport-photo
- transcribe-audio
- speech-say
- deliver-media
- reaction-gif
- get-media-info
- watch-video
- weather-get
- url-fetch
- web-content-fetch
- web-act
- web-drive
- consult-gemmy
- cascade-bridge
- code-analyze
- problem-analyze
- zoltraak-scan
- self-bucket-list
GrowthRecent Active Work
- pilot
- diagnose
- self-diagnose
- app-launch
- monitor-control
- power-usage
- library-engine
- library-acquire
- library-operate
- library-sync-plan
- avatar
- avatar-control
- avatar-overlay
- avatar-pilot
- screen-record
- audio-cleaning
- comic-analyzer
- read-manhwa
- record-inner-audio
- usb-vault-backup
- media-validate
- grok-image-gen
- grok-video-gen
Build Factory
From three lanes to a pilot bench.
Aiko's first long-running build system used three Gemini lanes: Architect, Executor, and Auditor. The current runner is broader: it can hand bounded work to Codex, Claude, Antigravity / Gemini, or Cursor, while Aiko stays the front door and keeps each friend or pilot session separate from her own memory.
v1 Lane · Planner
Architect
The only agent that talks to me. Source code is read-only. Owns the canonical task plan (CSOT.md). Drives BRIEFING → ALIGNING → REVIEW phases.
v1 Lane · Implementer
Executor
Pure code implementer. Never plans, never talks to me. Full CRUD on source + tests. Stages clean commits in the [AIKO:TASK] format. Never touches node_modules, .aiko/, or Aiko.md.
v1 Lane · Gatekeeper
Auditor
Verification gatekeeper. Read-only on source code, full write on tests. Never rewrites production code. Routes the FSM forward, back, or to BLOCKED.
- $build
- BRIEFING
- CODING
- AUDITING
- ALIGNING
- REVIEW
- DONE
Live System
The first checkpoint.
This first recording captures the mobile surface as it existed at the end of April. It stays here as the original checkpoint, with the newer growth video below it.
Aiko v1 · Mobile Surface Walk-Through
Recorded 29 April 2026 — portrait capture from the React Native phone app over Tailscale.
- 9:16 portrait
- 09:22
Build Log · April to June 2026
Not a redesign. A growth record.
The project started as a private mobile interface to a Linux agent. Each checkpoint added another kind of continuity: a place to read, a body on the desktop, more capable hands, and model-independent identity.
-
A phone surface for one local agent
The first version connected an Expo / React Native app to the Linux daemon over Tailscale, with Telegram as a second way to reach the same conversation, tools, and memory.
- React Native
- WebSocket
- Tailscale
-
The Library grew beyond a tab
Source browsing, chapter downloads, comic and EPUB reading, resume state, narration, voice packs, and a navigable pixel room turned the app into a lived-in space rather than a chat client with extra screens.
- Suwayomi
- Offline reading
- Voice packs
- Pixel room
-
From a sprite to a desktop presence
The Agent Forge build connected Aiko's stage directions to a movable VRM stage, 79 retargeted motion clips, swappable skins, a read-only live chat tap, and Qwen3-TTS voice synchronization.
- Three.js
- VRM
- Kimodo motion
- Qwen3-TTS
-
PC and phone started sharing the same moment
The embodied experiment moved from isolated demos toward a live shared session: desktop presence, phone control, model and skin switching, and the motion pipeline shown as one evolving system.
- PC + phone
- Live session
- Skins
- Motion research
-
Identity separated from the brain underneath
Aiko now has 88 active skills and can route between local Ollama, OpenRouter / MiniMax, Moonshot / Kimi, and CLI agents. Provider tags and isolated friend sessions make the system easier to inspect without fragmenting the one identity the user talks to.
- 88 skills
- Kimi
- CLI pilots
- Provider attribution
Aiko, then and becoming.
A 93-second build diary: the pixel Library room, the first humanoid / VRM experiments, one live PC-and-phone session, swappable model skins, and the motion research that is shaping the next body.
- 9:16 portrait
- 01:33
- H.264 + AAC