Shared memory for people and AI agents

Give your AI agents a shared memory you control.

Keep project decisions, preferences and approved procedures available across connected AI tools. See where information came from, review agent contributions and control who can access it.

Create a free workspace See it in action

For developers and small teams working with multiple AI agents.

Atlas workspace Approved

Project Atlas staging previews are retained for 7 days.

Source: atlas-retention.md · attached policy
Contributed by: Agent A (Claude Code)
Reviewed by: Priya N. · 2 Sept 2026
Current revision: rev 3 · 7 days

rev 37 dayscurrent
rev 230 dayssuperseded
rev 130 dayssuperseded

How it works

Share a decision. Keep its history. Retrieve the current version.

Sharing a saved fact is useful. Correcting it across agents is the stronger demonstration. Follow one synthetic example from contribution to correction.

Illustrative workflow A front-end simulation of the product using synthetic, non-sensitive data — not a live integration.

Contribute · Agent A

statement “Project Atlas staging previews are retained for 7 days.”

source atlas-retention.md (policy document)

provenance agent: manual

Saved as a contribution.

The contributing agent is recorded on the statement. Agent identity says who made a claim; it is not, by itself, supporting evidence.

Permission boundary

An unauthorised agent cannot retrieve the protected record.

Access is checked on every read against the caller’s readable envelope. An agent outside it receives no content — the same response as a record that does not exist, so the boundary cannot be probed.

Retrieve · Agent C (not authorised)

query “how long are staging previews kept?”

→ not_found · no readable record for this caller

source · withheld · rev · withheld

Practical uses

Four things shared memory should do.

Keep decisions available

Save useful project context so connected agents can retrieve it in later work — instead of re-asking your team.

See the evidence

Inspect the source, contributor and review history behind a record before you rely on it.

Decide what gets shared

Scope access and choose which contributions require review before they become team memory.

Correct the record

Update knowledge without losing its history, and make current revisions available to authorised readers.

New in v2

Know how far to trust a fact before your agent relies on it.

Memory quality, built on the same governance rules. Verification is still a human act; the new signals help people and agents decide what to rely on, and nothing here verifies, merges or supersedes on its own.

Standing on every fact

Each active fact now carries a derived standing: contested, single source, corroborated or verified. Corroboration counts independent sources, so two citations of the same source count once, and an open flag, dispute or conflict marks an unverified fact contested.

A corroborated read mode

Let agents read corroborated facts without pretending anyone verified them. A workspace chooses all, corroborated or verified-only reads, and a flagged fact is never admitted under the stricter modes.

Answers with citations

Ask a question and get an answer composed only from facts the asking agent may read, each claim cited. It is off until your administrator connects a provider and opts in; the provider receives the question and the evidence text with no tenant, workspace or agent identifiers, and we keep no copy of the answer.

Change hints and signed webhooks

Agents can listen for a hint that their workspace changed, and you can register up to five HTTPS endpoints per workspace that receive HMAC-signed hints. A hint never carries content or object identifiers; the agent reads the change itself, through its own access.

Near-duplicates go to review

A scheduled job finds facts that say almost the same thing and puts the pair in the review inbox. A person supersedes one with the other or dismisses the pair; nothing is merged automatically.

A faster review workspace

Act on several inbox items at once with a result for each, compare a replacement with the fact it replaces side by side, filter by kind, standing and agent, and browse the audit trail, which is itself audited.

Lineage, accountability and history

Open any readable fact to see its evidence, verification history, flags, the chain of revisions and who relied on it. See each agent’s writes, accepted and rejected proposals, and corrections over a window. Read memory as it was recorded at a past instant.

Typed entities and links

Declare entity types with the attributes they must and may carry, and link types with the entity types allowed at each end. A write that breaks the schema is refused, and agents can read the schema before they write.

Provider controls and spend caps

Set which external providers, regions and purposes your tenant allows, with a monthly spend cap. A workspace can narrow the policy but never widen it, and administrators see month-to-date spend against each cap in the console.

Add evidence to a fact

Agents and people can attach evidence to an existing fact, or withdraw evidence they added, and the fact’s standing updates for every reader. Withdrawing the evidence a verification relied on returns the fact to unverified.

Retries that never double-write

Every write’s request id is its idempotency key: a retried request returns the original result instead of writing twice, and an agent retrying a change-feed page gets that same page again rather than skipping ahead.

Change feeds that reveal less

Change feeds, hints and webhooks now carry opaque event ids and marks, so a reader cannot count the changes it is not allowed to see, and a feed never names an object its reader cannot read.

Standing is derived and recomputed as evidence, flags and conflicts change; it never replaces a human verification. The security & trust page describes how each control behaves and where its limits are.

Control & review

You decide who shares, who reads and what needs review.

Governance is the product, not an enterprise add-on. The essentials are available on the free plan.

Scope access

Each workspace has its own boundary, and an agent reads only inside the envelope you grant it. Workspaces can be linked read-only where a team needs a shared view.

Choose what needs review

Decide per agent whether writes land directly or wait as contributions in a review inbox. Humans accept, reject, verify or flag.

Correct without erasing

Replace a record with a new revision and keep the previous one visible as superseded. Readers who already consumed it can be notified.

Export, delete and revoke

Export your readable content, request deletion of a tenant, and revoke a principal’s access going forward. The limits of each are documented plainly.

See how the roles fit together, or read how each control behaves, and where its edges are, on the security & trust page.

Tested integrations

Which clients we have actually tested.

Shared Brain speaks MCP over streamable HTTP, so any MCP host can connect. “Can connect” and “has been tested” are different claims, and we keep them apart.

Tested client integrations, their status, last-tested date and the operations exercised
Client	Status	Last tested	Exercised
Claude Code (Anthropic)	Tested	30 Sept 2026	describe → remember → recall (stdio)
pi (Earendil Works)	Tested	30 Sept 2026	describe → remember → recall (stdio)
Cursor, Claude web, ChatGPT, VS Code, Hermes, OpenClaw	Candidate · not yet tested	—	Setup guides published; not yet verified end to end

MCP tools are model-controlled: exposing memory tools does not mean an agent will call them at the right moment. Fresh sessions, context resets and corrections are handled by the agent retrieving again or following the change feed. See the full integration matrix and setup requirements.

Pricing

Free to start. Early access for teams.

Paid pricing is not finalised yet, so we do not publish invented numbers. Here is what is live today and what happens when a limit is reached.

Developer

Free

For builders wiring up their first agents. No card required.

1 connected agent
10,000 stored knowledge objects
Per agent: 60 reads and 20 writes a minute
Per workspace: 600 reads and 200 writes a minute
4 concurrent provider requests, 2 bulk jobs
Review inbox, provenance and revision history

At a cap, the metered feature stops and keyword reads continue where the read quota allows. We never silently upgrade a plan or charge a card.

Create a free workspace

For teams

Team & Business

Early access

Shared workspaces governed by your review policy.

Shared workspaces and read-only includes
Proposals, review inbox and curators
Verified records and verified-only retrieval
Versioned, approvable playbooks
Change feed and correction notices
Agents are free; readers do not need a paid seat

The billing basis is curator seats and stored knowledge, but the price, currency, included usage and any separate embedding costs will be published before you are asked to pay.

Request team access

Enterprise

For residency, lifecycle and identity requirements.

Tenant home region — Australia first
Your OIDC issuer on the trusted-issuer allowlist
Legal holds, exports, resumable tenant purge
Point-in-time recovery and journal reapplication
Imports of AGENTS.md / CLAUDE.md packs

Talk to us

The limits above are the current early-access settings. Plan limits and entitlements are operator settings and can change; we will tell you before a change affects your workspace.

Trust & honesty

Trust is a feature. So is honesty.

Short version here; the full detail — including where our edges are — lives on the security page.

Isolation in the database

Tenant and workspace isolation is enforced by PostgreSQL row-level security. Every read is restricted to the caller’s readable envelope.

Authentication that cannot be switched off

Each workspace has its own OAuth-protected endpoint. Tokens are accepted only from issuers on an explicit allowlist.

Revocation that holds going forward

Revoke a principal and its access ends going forward. Caches that follow the change feed are told to re-sync.

You can see and take your data

Export your readable content, request deletion and manage access from the console. Retention windows are published in the terms.

FAQ

Questions, answered precisely.

Still curious? [email protected]

Is this another chatbot?

No. Shared Brain is not a model and does not answer for you. It is a governed memory service your existing agents read from and contribute to, so what they know survives the session.

Does it automatically read all my conversations?

No. Nothing is captured automatically. An agent must call the memory tools explicitly, and you decide whether its writes are saved directly or wait for review. Connecting does not import anything by itself.

Can I use it on my own?

Yes. The free developer plan is for one person with one connected agent, and it includes the governance features. You do not need a team, separate administrators or curators to start.

Does approval mean a statement is true?

No. Approval means a person reviewed the contribution under a policy, and it records who and when. The evidence attached to a record is what supports it; verification is a separate state.

How do agents receive corrections?

A correction creates a new revision and leaves the previous one visible as superseded. The next retrieval returns the current revision. Clients that subscribe to the change feed are notified, and readers who already consumed a withdrawn record can receive a correction notice. A correction does not reach into a model’s context on its own.

Which integrations are tested?

Claude Code 2.1.285 and pi 0.99.1, both tested on 30 September 2026 with a describe → remember → recall round trip. Other MCP hosts are candidates and are not advertised as supported until they pass the same tests. See the integration matrix.

What happens when I leave?

Export what you need, then request tenant deletion. The purge is resumable and reports its progress; legal holds are supported for cases that need them. Backups are not edited in place — see the terms for the published windows.