GitHub 中文

Open-source Personal Agent State Layer

Memory isn't enough.

Trusted state for every AI.
Track goals, constraints, and progress as they change. Know what's confirmed, uncertain, or conflicted before the next action.

Try the demo ▼
Change it yourselfIn-browser demo. No model involved.
codexChat
>
▸▸ Auto-sync on
Problem

Problem & Aha

Your AI remembers. But is it still right?

Even with one agent in one session: a constraint changes, a tool result is never confirmed, the evidence conflicts, and it carries on with the old assumption.

Illustrative scenarios of what can go wrong when an agent mistakes a stale assumption for current state. Not every memory product or agent makes these mistakes.

Remembering isn't the same as staying current.

Help agents act on trusted, traceable state—not stale context.

Context rot can make important changes harder to use in long tasks; state drift also includes uncertain tool results and unresolved conflicts. Track what is confirmed, unknown, or conflicted before acting. The goal is a current state view with evidence behind it, not a guarantee that every step is right.

Memory1 / 5

Memory and context are the input.

It's all in the context. Nothing says what still holds.

In one long-running task, instructions, revisions and tool results pile up in the context. Every line is there. None is marked as still in effect.

TASK CONTEXTScattered
“Everyone can export orders”“Make it admins only”“Deploy v2”“Request timed out”“Tracker: complete”
? Which still count
State View2 / 5

Trusted State View is the product.

One Trusted State View

Goals, constraints, plans, progress: each has a current value, with its source and how far it has been confirmed. An inference never passes as confirmed.

SAVESLOT · Order exportv1
GoalOrder export APIConfirmed
Export accessEveryoneConfirmed
Endpoint/v2/orders/exportInferred
v2 deployNot deployedConfirmed
Task #7In progressConfirmed
Updates3 / 5

When constraints change, the state follows.

When a constraint changes, the state follows.

Say it once. The old constraint is marked “superseded” and kept in history; what the agent reads next is the new constraint and where it came from.

Old versions are never deleted; view or roll back any time. The above is the target experience, to be validated in the product.

SAVESLOT · Order exportv2

“Only admins may export orders.”

Export accessEveryoneSuperseded
Export accessAdmins only★ CONFIRMED
GoalOrder export APIConfirmed
Endpoint/v2/orders/exportInferred
v2 deployNot deployedConfirmed
Task #7In progressConfirmed
Validation4 / 5

Unconfirmed is not done.

Unconfirmed is not done.

A deploy request times out: the result is Unknown. Two systems disagree: the status is Conflicted. The agent verifies or asks you first, instead of filling the gap or quietly picking one.

SAVESLOT · Order exportv3
Export accessAdmins onlyConfirmed
v2 deployRequest timed outUnknown
Task #7Sources disagreeConflicted
Endpoint/v2/orders/exportInferred
Same session · before the next action

The export endpoint is done, admin-only. The v2 deploy isn't confirmed yet, so I'll verify it with the deployment system first. Task #7 has two conflicting reports and needs your call.Read v3

Continuity5 / 5

Cross-session continuity comes next.

A new session or another agent can pick it up.

This comes next: a new session, or another agent you've authorized, reads the current version of this view and the reasons behind each change. Being able to read it doesn't grant write access to the source system.

Illustrative flow of the target product experience. What ships at launch is what counts.

YOUR SAVESLOTv4
Export accessAdmins onlyConfirmed
v2 deployDeployed★ CONFIRMED
Task #7Sources disagreeConflicted
Endpoint/v2/orders/exportInferred
CodexWrote v4Claude CodeAuthorized · readChatGPTAuthorized · readCursorNot authorized

Trusted State View

What makes state trustworthy?

A Trusted State View is not the one true record. It keeps the source, evidence, effective time, version and permissions for each piece of state, and marks how far it has been confirmed. Confirmed only means an appropriate authoritative source confirmed it, not that it is absolutely true. Business facts and write authority stay with the source system.

The four trust markers and four capabilities are product goals. All are planned; none is shipped or verified yet. At launch we'll publish each one's status with reproducible tests.

OK

Confirmed

Confirmed by an appropriate authoritative source, with who confirmed it, on what basis, and when. It does not mean absolute truth.

~

Inferred

Inferred from conversation or tool evidence, not yet confirmed by an authoritative source. It is never silently promoted to Confirmed.

?

Unknown

The information is missing or can't be verified. The agent must admit the uncertainty instead of filling it in.

!

Conflicted

Sources or evidence disagree and it isn't resolved. No unilateral ruling: it is handled by permissions, source priority and a confirmation step.

Every piece of state also carries time, validity period, version, source and evidence, and permissions. “Expired” can sit alongside any of the four markers.

Four capabilities

Keep the view maintained, and hand it over before the agent decides.

01

State Tracking

Track goals, constraints, plans, decisions, commitments and key progress, so changing state becomes an explicit, maintainable object.

Planned
02

State Updates

Update state from new instructions, tool results and outside changes, keeping versions, the source of each update, and superseded / effective relationships.

Planned
03

State Validation

Mark state as Confirmed, Inferred, Unknown or Conflicted, detect what has expired or lapsed, and show the evidence, the time, and what still needs confirming.

Planned
04

State Retrieval

Before the agent acts, return a compact, relevant view: current values, trust markers, sources, and any conflict notes it needs.

Planned

Your data stays yours

Your state doesn't belong to any AI platform. The goal: visible, portable, deletable.

{}

Open-source core (planned)

We plan to publish the code, with full core features in the self-hosted version.

↓

No platform lock-in

State is independent of any AI platform. Uninstalling a plugin doesn't lose a save; export and delete are planned to be free on every plan.

◐

Least privilege

Only AIs you authorize can read or write. Each plugin declares what it touches and whether it goes online. Revoke any time. Reading state doesn't grant the right to change the source system.

✓

Public audits (planned)

We plan to publish SECURITY.md and a threat model. There are no third-party audit findings yet.

How we measure it

How we measure it

“Remembers it right” can't rest on a slogan. These are the six metrics we will measure. They are targets to validate, not results. We'll also measure what a separate state layer costs in latency, context overhead and user confirmations.

■ We'll publish the method and reproducible results at launch. Until then, no numbers here.

At launch

Same-session state drift rate

State drift

Within one long-running task, how often the state an agent acts on is no longer the value in effect.

At launch

False tool-step completion rate

Tool results

How often a tool call that failed, timed out or only partly succeeded gets marked as done.

At launch

Lapsed-constraint misuse rate

Misuse

How often an agent acts on a constraint or plan that has already been superseded.

At launch

Trust-marker classification quality

Classification

Whether Confirmed, Inferred, Unknown and Conflicted are assigned correctly.

At launch

Evidence traceability coverage

Provenance

How much of the state can say where it came from and why it changed.

At launch

Cross-session and cross-agent continuity

Continuity

After a new session or a switch to another authorized agent, can work continue from the current version?

Plugins

One core. One plugin per job.

Plugins share one system of versions, permissions, provenance and rollback. Add a second one without learning it again.

01

Project Keeper

At launch

Goals, constraints, decisions and progress in long-running tasks, carried across sessions and coding agents too.

02

Body Log

At launch

Meals, workouts, weight and a weekly recap. No medical diagnosis.

03

Goal / Decision Tracker

Example

Goals, decisions and why they changed. A demo of how to build a new plugin.

--

Build your own plugin

Open

Job search, budget, study, family… build your own with the Plugin Starter. No core changes needed.

Waitlist · step 2

What should your AI always remember?

Optional; skip it if you like. We read every answer and use them to decide what to build first.

Used only to tell you when we launch and to understand what you need. Privacy notes

PREVIEW · after launch, you hand your AI this prompt to connect

Planned pricing

Pick your plan

■ Planned pricing. May change before launch.

■ Export, delete, correct and manage permissions: planned to be free on every plan

Click a card or use ← → to pick a plan, then join the waitlist

FAQ

Questions you might have

Didn't find your answer? , I read every one. Or ask on Discord directly.

Get on the list

Don't let your AI act on stale assumptions

What we're building toward: you state a change once, unconfirmed results are marked unknown, and your AI reads evidence-backed current state before it acts.Leave your email and we'll tell you the moment we launch.

Back to the demo
AIDEN · FOUNDER

Me again. You made it to the end, which is more patience than my own AI has. Tell me what you didn't get, where you think I'm bragging, or what made you want to swear. I read every message myself, and I remember them right. To rant in person, find me on Discord.

GET saveslot.ai/state.json200 OK

Same site. This is what an agent reads.