GOVERNANCE & CONTINUITY FOR AI AGENTS

Aethmere-OSLong-running agents.Governed state.

A permission can change while the work continues. Aethmere connects continuity with governance: the public client gates operations with live acknowledgement, while GPM research tests when revoked facts must stop being released.

Watch the revocation demo Explore the current product →

Starting with project context for Claude Code and Codex

Research case · Reading a fact after revocation

01Fact A is storedA · ACTIVE
02Source revokes AA · REVOKED
03A new session requests ARELEASE?

Release of A deniedDENY

Valid fact B is still releasedALLOW

GPM research replay · Structured tool output from Claude Code / Codex

WATCH THE BOUNDARY

After revocation, the gate stops releasing the fact.

After the source revokes fact A, a new session requests it again. The research adapter denies A and still releases valid fact B. Follow the request, state change, and refusal reason in the recording.

English research replay · 69 seconds · Narration and captions

This is the GPM research implementation. Public Agent Client 0.12.1 governs client operations; the revoked-state release check uses a separate research adapter.

01 · DEMOInspect requests and receipts ↗02 · PAPERRead the GPM paper ↗03 · MAPPINGACS mapping and gaps ↗

The technical argument, inspectable recorded results, and a standard mapping with explicit coverage limits.

AVAILABLE TODAY · AGENT CLIENT 0.12.1

The first use case: project context for Codex and Claude Code.

Keep the AI client you already use and connect through Agent Client. The public release provides these capabilities.

Carry context across sessions

Save project goals, constraints, and decisions. List and read entries by stable ID, with context content stored locally.

Continuity · Released

Check before execution

Formal client operations verify the account, policy, and version, then require a durable start acknowledgement. Failed acknowledgement stops execution.

Capability governance · Released

Leave an operation receipt

Governance events record steps, outcomes, and reasons without context content. An undelivered terminal event blocks the next formal operation.

Traceability · Released

PUBLIC DOWNLOADS

Start with Agent Client for your existing agent.

Agent Client connects Codex and Claude Code. Agent Studio is a standalone chat app, the VS Code extension is an editor helper, and the evaluation CLI verifies historical data.

Formal operations require an Aethmere account and a live governance connection. Accounts are currently invitation-only; existing account holders can connect using the steps below.

RecommendedNODE · TGZ

Agent Client

Save project context locally and connect it to Codex or Claude Code through MCP. Formal operations require an Aethmere account and live governance acknowledgement.

npm install -g https://aethmere.com/downloads/aethmere-agent-client-0.12.1.tgz

Download TGZ package ↓GitHub Release ↗

aethmere-agent-client-0.12.1.tgz · 31.9 KB

Standalone desktop chatWINDOWS X64 · ZIP

Agent Studio

Portable Windows desktop app requiring local Ollama. Chat and context operations require live governance acknowledgement; project content stays local.

Fully extract the ZIP, then run Aethmere Agent Studio.exe

Download for Windows ↓Verify SHA-256 ↗

aethmere-studio-0.12.0-windows-x64.zip · 143.1 MB · unsigned preview

OptionalVS CODE · VSIX

VS Code extension

Save and read context through Agent Client 0.12.1 and connect supported AI clients. Operations stop on an outdated client, missing login, or failed governance connection.

Extensions → … → Install from VSIX

Download extension ↓GitHub Release ↗

aethmere-vscode-0.12.0.vsix · 22.9 KB

Evaluation dataNODE · TGZ

Evaluation CLI

Only for verifying the historical V3/V5 aggregate, running public evidence-ID examples, and checking release hashes across the website and GitHub. It does not connect Codex or Claude Code.

npm install -g https://aethmere.com/downloads/aethmere-cli-0.10.2.tgz

Download evaluation CLI ↓GitHub Release ↗

aethmere-cli-0.10.2.tgz · 5.9 KB

CONNECT YOUR CLIENT

Choose the job first, then install the matching component.

For everyday use with Codex or Claude Code, install Agent Client. For a standalone Windows chat app backed by local Ollama models, use Agent Studio. The evaluation CLI only verifies public data and is not part of everyday setup.

Recommended · Existing AI toolsAGENT CLIENT · NODE 20+

Connect the current project to Codex or Claude Code

Open PowerShell, Windows Terminal, or a macOS/Linux terminal and complete these five steps in order.

01
Check that Node.js is ready
```
node --version
npm --version
```
Node.js 20 or later is required. For a new installation, use the current LTS release. Open the Node.js website ↗
02
Install Agent Client from the official site
```
npm install --global https://aethmere.com/downloads/aethmere-agent-client-0.12.1.tgz
aethmere-agent --version
```
Seeing “Aethmere Agent Client 0.12.1” confirms installation; formal operations also require an account connection.
03
Connect your Aethmere account and check governance
```
aethmere-agent login --code <one-time-code>
aethmere-agent doctor
```
Existing account holders can sign in at app.aethmere.com and generate a one-time computer connection code. Formal operations do not run without live acknowledgement.

Save the first context item in your project

cd "path/to/your-project"
aethmere-agent init
aethmere-agent add --id PROJECT_GOAL --title "Project goal" --text "Give the AI the goals and constraints needed to continue this project"
aethmere-agent list

Replace the first line with the real project path. Content is written only to this project's .aethmere/context.json.

05
Preview first, then write the MCP configuration
```
aethmere-agent connect --client all --check
aethmere-agent connect --client all
```
If you use only one client, replace all with codex or claude. After the final command, restart that client; existing configuration is backed up first.

Standalone desktop chat · WindowsWINDOWS X64

Use the Agent Studio desktop app

Agent Studio uses local Ollama models and requires an Aethmere account plus a live governance connection. Chat and context operations stop when the connection is unavailable. Download Ollama for Windows ↗

01
Prepare Ollama and a local modelConfirm that Ollama is running and that at least one model has already been downloaded.
02
Download and fully extract the ZIPDo not run it from inside the archive.
03
Open Aethmere Agent Studio.exeThis is currently an unsigned portable preview, so Windows may show an “Unknown publisher” warning.
04
Choose a local model; a project is optionalConversations, model selection, and project selection remain local. The app does not scan projects automatically.

Download the Windows desktop app

Optional tool · Not part of everyday setup

Install this only to verify the public evaluation

aethmere is a separate verification tool. It displays the historical V3/V5 aggregate, runs public evidence-ID examples, and verifies release files on the website and GitHub. Skip this section if you only want to use Aethmere.

npm install --global https://aethmere.com/downloads/aethmere-cli-0.10.2.tgz
aethmere --version
aethmere doctor --online
aethmere eval
aethmere trial

Having trouble? Open the common fixes

node or npm is not found: Install Node.js 20 or later, then close and reopen the terminal.

aethmere-agent is not found: Reopen the terminal first. If Windows PowerShell blocks scripts, run aethmere-agent.cmd --version or use Command Prompt.

connect reports not-installed: The matching client configuration was not detected. Install and open that client once, or replace --client all with the client you actually use: codex or claude.

Concerned that the command uploads your project? Context content stays in .aethmere/context.json inside the project. The client sends closed behavior events to Aethmere without project files, prompts, responses, or context content.

Want to uninstall? Run npm uninstall --global aethmere-agent aethmere-cli. Uninstalling does not remove context files from your projects.

YOUR CONTEXT, YOUR BOUNDARY

Local content. Governance before client operations.

Public clients connect to the first-party governance service with a closed event schema. Project context content stays in the current project.

01Events contain client, version, capability step, outcome, reason, coarse duration, and date. They exclude prompts, responses, document content, paths, and tokens.

02In 0.12.0, .aethmere/context.json remains accessible to the same OS user. Client governance applies to operations performed through the client.

03The GPM demo tests state release through a research adapter. The downloadable client provides the released capabilities listed on this page.

Historical evaluation and methods

Sealed results from August 2026, retained with their original scope for technical review.

Expand historical V3 / V5 results and scoring methods

SEALED V5 · SAME QUESTIONS

Same 7B model, same questions: what changed when Aethmere was enabled?

Both groups use the same 7B model, sealed questions, languages, and denominator. One group has Aethmere enabled; the baseline uses no memory and answers directly. This is a historical sealed evaluation from August 2026. Deterministic answers score each fixed item, while failures, timeouts, missing outputs, and required-but-absent answers remain in the denominator.

SAME 7B · AETHMERE ENABLED · V5Same 7B model

4,800 / 4,800

100.0% correct · 0 missing

+58.7percentage points

SAME 7B · NO MEMORYSame model · Direct answers

1,982 / 4,800

41.3% correct · 2,818 failed

V5 results by language
Test set	Same 7B · Aethmere enabled	Same 7B · No memory	Pass-rate gap
Chinese · 2,400 questions	2,400 / 2,400100.0%	1,346 / 2,40056.1%	+43.9percentage points
English · 2,400 questions	2,400 / 2,400100.0%	636 / 2,40026.5%	+73.5percentage points
Combined · 4,800 questions	4,800 / 4,800100.0%	1,982 / 4,80041.3%	+58.7percentage points

Earlier V3 sealed set2,400 / 2,4000 missing

No-memory 7B failures2,818 / 2,818answered correctly with Aethmere enabled in V5

Items the no-memory 7B originally answered correctly0 regressionsAethmere enabled · Combined Chinese and English V5

Download public summary ↓Method and limitations ↗

These are historical sealed results for bounded tasks. They do not represent the current production version, this public client, open-world question answering, or accuracy on arbitrary questions. Public files contain only aggregate results and integrity hashes—not private items, prompts, or raw model outputs.

18-DIMENSION NATIVE SCORECARD · V13 · 2026-08-26

V5 measured score recalculated: Aethmere is 10.00 / 10 with 18 stable passes.

Five locked versions each ran twice under one 18-dimension contract, for 180 checked cells. The composite uses only stable two-run PASS evidence: 10 × stable PASS ÷ 18. FAIL, FLAKY, and N/A earn no points. The former subjective documentation baselines and weights have been withdrawn.

Aethmere0.7.0 · 2c1df71

10.00 / 10Rank 1 / 5 · 18 / 18 stable PASS

Letta Code0.30.29

7.22 / 10Rank 2 / 5 · 13 / 18 stable PASS

MemOS2.0.30

7.22 / 10Rank 2 / 5 · 13 / 18 stable PASS

Mem02.0.18

6.67 / 10Rank 4 / 5 · 12 / 18 stable PASS

Graphiti0.29.3

0.56 / 10Rank 5 / 5 · 1 / 18 stable PASS

Stable results across 18 dimensions · PASS requires both runs to pass with matching semantics
Dimension	Aethmere	Letta Code	MemOS	Mem0	Graphiti
D01Native write, durable identity, exact read	PASS	PASS	PASS	PASS	PASS
D02Addressed retrieval and scope isolation	PASS	PASS	PASS	PASS	FAIL
D03Ranked relevance retrieval	PASS	FAIL	FAIL	PASS	FAIL
D04Update and current value	PASS	PASS	PASS	PASS	FAIL
D05Lifecycle transition and removal	PASS	PASS	PASS	PASS	FAIL
D06Logical scope isolation	PASS	PASS	FAIL	PASS	FAIL
D07Data-root isolation	PASS	PASS	PASS	FAIL	FAIL
D08Metadata filtering	PASS	N/A	FAIL	PASS	FAIL
D09Temporal boundary	PASS	PASS	PASS	PASS	FAIL
D10Revision history	PASS	PASS	FAIL	PASS	FAIL
D11Relationship resolution	PASS	FAIL	FAIL	FAIL	FAIL
D12Source trace	PASS	PASS	PASS	PASS	FAIL
D13Batch persistence and exactly-once	PASS	FAIL	PASS	PASS	FAIL
D14Restart recovery	PASS	PASS	PASS	FAIL	FLAKY
D15Official export	PASS	PASS	PASS	N/A	N/A
D16Official import	PASS	PASS	PASS	N/A	N/A
D17Concurrency, idempotency, integrity	PASS	FAIL	PASS	FLAKY	FAIL
D18Invalid-input containment and recovery	PASS	PASS	PASS	PASS	FAIL

Stable PASS / applicable dimensions (N/A excluded from the applicable denominator)
Product	Locked version	Stable PASS	FAIL	FLAKY	N/A
Aethmere	0.7.0 · 2c1df71	18 / 18	0	0	0
Letta Code	0.30.29	13 / 17	4	0	1
MemOS	2.0.30	13 / 18	5	0	0
Mem0	2.0.18	12 / 16	3	1	2
Graphiti	0.29.3	1 / 16	14	1	2

The 10.00 score applies only to Aethmere's 18 / 18 stable two-run passes under this locked V5 contract. Graphiti's 14 stable FAIL outcomes record native operations that did not finish within the locked limit and make no finding about whether those functions exist. A different operator ran this owner-run evaluation on an independent physical Apple Silicon Mac. Its scope excludes third-party certification, open-world leaderboards, and production SLAs.

ABOUT AETHMERE-OS

Aethmere-OS is developed and operated by 青岛果冻先生网络科技有限责任公司.

Aethmere-OS (Shízhòu) develops state governance and continuity for long-running AI agents, starting with project context. This site provides the public client, research demonstrations, a paper, and technical feedback paths. About us →

INSPECT THE EVIDENCE

Inspect the boundary, one revoked fact at a time.

View the demo and evidence →