More Than a Chat Window

AILIS is designed to become a personal AI that actually lives on your desktop. She has a visible 3D character, voice, expressions, and long-term memory, backed by an Agent Runtime that can research, read files, write code, organize content, and operate computer tools.

Talk to AILIS naturally, like a companion. When there is work to do, she can understand approved screen and file context, choose the right tools, complete the task, and remember the preferences that matter next time.

Core Experience

Visible A VRM desktop character with expressions, motions, lip sync, and dialogue bubbles. AI no longer has to feel like an empty text box.	Conversational Voice input and natural speech output are available alongside quiet, fast text interaction.	Context-aware With permission, AILIS can understand screens, windows, captured regions, and local files without making you repeat the context.
Capable Search, code, files, web, email, and computer actions share one auditable tool execution path.	Memorable Long-term memory keeps useful preferences, project background, and relationship context for better collaboration.	Controllable Important tool actions enter approval and audit flows, so users know what the system plans to do and what it has done.

How AILIS Works

1. Describe	2. Understand	3. Execute	4. Remember
Explain the goal naturally	Read approved screen and file context	Use search, code, file, and computer tools	Keep useful preferences and project background

Evaluation Results

Agents' Last Exam · Same-Model Comparison

GPT-6 Luna · xhigh · 98 matched tasks · October 2026

ALE metric	AILIS · previous	AILIS · evaluated candidate	Codex
Mean task score	42.63%	49.35%	48.53%
Full / partial / zero credit	20 / 37 / 41	23 / 43 / 32	24 / 40 / 34
Logical input tokens	270.99M	208.49M	488.25M
Uncached input tokens	16.76M	28.72M	12.25M
Output tokens	4.52M	4.27M	3.34M
Input cache rate	93.81%	86.22%	97.49%
Mean task time	2,237.4s	1,517.9s	1,694.6s

The Docker manifest contains 99 tasks; 98 have valid paired results. One unavailable API-dependent verifier is excluded from both systems. Scores include partial credit; timing uses 93 tasks with matched records, including network/setup time. These are local 9,000s-budget runs, not the official 900s leaderboard protocol. The +0.82 pp lead over Codex is a single-run observation. Protocol, per-task scores and resource accounting.

The evaluated snapshot predates v1.4.9's context-window correction; no new evaluation was launched for this release.

Earlier Terminal & General-Agent Results

End-to-end Agent evaluation · 2026-09-07 · scores in %. AILIS is highlighted; columns identify both the system and its model / reasoning setting.

[1] Terminal-Bench: 65 passes / 87 valid tasks, with 2 of 89 unresolved. Codex is the official archived mean across 89 tasks, with five trials per task. [2] GAIA: reviewed answers on 122 completed tasks, with 43 of 165 unresolved; AILIS uses Max / 3000s, the Codex references use Medium / 600s. These are different-budget frozen-run results, not full-suite certification of the current release.

Text table · scores, sample counts, and missing values

Capability / Benchmark	AILIS Luna Max	Codex Luna Max	Codex Luna Medium	Codex GPT-5.5 Medium
Coding
Agentic terminal coding Terminal-Bench 2.1 [1]	74.71 65 / 87	75.73 89 tasks · 5-trial mean · ± 1.32 SE	—	—
Agent
General-purpose tasks GAIA · Overall [2]	77.05 94 / 122	—	67.21 82 / 122	79.51 97 / 122

Scores are percentages. Bold indicates the highest reported value in a row, not a controlled ranking. — means no corresponding evidence, not failure or zero.

Memory & Humanlike Companion

LongMemEval-S and LoCoMo use the August 5 memory baseline; Humanlike uses the July 20 internal longitudinal summary. These are separate historical evaluations, not Luna Max results from the task-execution table or reruns of the current unified Agent. LoCoMo is token-F1, not a task pass rate; Humanlike is an internal rubric score, not an external benchmark ranking.

Text table · memory, stateful tools, and companion experience

Capability	Benchmark	AILIS historical result	Metric / coverage
Long-term question answering	LongMemEval-S	71.60%	QA accuracy · 358 / 500
Conversational memory	LoCoMo	24.69	token-F1 on a 0–100 scale · 1,986 questions
Humanlike companion	Humanlike · Longitudinal Agent Eval	78.46 / 100	171 judged checkpoints · 61.4% pass rate · 16 hard failures
Stateful tool use	ToolSandbox	71.51%	Frozen holdout mean
Personalized memory	PersonaMem Balanced-140	65.71%	92 / 140

Humanlike checkpoints come from 30-day companion scenarios, not a 30-day real-user study. See the full scorecard for retrieval recall, latency, experience dimensions and archived sources. No matched Codex results are available for these three evaluations.

Full scorecard, efficiency, historical comparisons, and evaluation protocols →

What Works Today

• A resident VRM character, chat window, and control panel on Windows
• Realtime interaction through text, voice, expressions, and motion
• Permission-aware screen, window, file, and code context
• Search, web, code, file, email, and computer-operation tools
• Long-term memory for preferences, projects, and relationship context
• Approval, evidence, and recovery paths for consequential tool actions
• Stronger reliability, caching, and recovery for long-horizon work
• A more complete realtime voice, cross-device, and plugin experience

Quick Start

v1.4.9 update

GPT-6 Luna now uses a 272k context window with compaction at 244.8k; unrecognized models default to 256k unless a window is explicitly configured. Existing prompts, tools and compaction algorithms are unchanged. Windows, Linux and both Mac architectures retain bundled offline speech recognition and wake-word support. See the release notes and v1.4.9 downloads. v1.4.8 remains available for rollback.

Use AILIS

Download the desktop build from Releases, or meet AILIS first through the web experience.

Develop Locally

pnpm install
pnpm desktop:dev

Desktop builds, voice, validation, the optional backend, and packaging are documented in the Getting Started guide.

Direction

AILIS is neither a roleplay chat app with no execution ability nor a terminal wrapped in an avatar. The project brings three ideas together:

1. A digital companion with presence: conversation, voice, expression, relationships, and long-term memory.
2. A reliable personal Agent: contextual understanding, general tools, and long-horizon task execution.
3. An understandable, controllable execution system: approved actions, traceable progress, and recoverable failures.

Learn More

Install and Configure · Documentation Center · Complete Evaluation Scorecard

Community

If AILIS is useful to you, star the repository to follow its progress. Bug reports, workflow ideas, and focused pull requests are welcome through Issues and the contribution guide.

Privacy and Control

AILIS is built for personal desktop use. Visual context requires permission, actions that affect files, apps, accounts, or external services enter an approval flow, and local memory and runtime state remain on the user's machine by default. Only context needed for the current request is sent to the configured model service.

License

AILIS source code is released under the MIT License. Some third-party models, motions, voice assets, and character resources may use their own licenses.