Context does not compound
People re-brief each assistant, so yesterday’s context has to be explained again.
Shared memory for teams · working name
Your team's AI should share what it learns.
Super orchestrator of the AI agents
Continuum is a working name for a vendor-neutral, governed memory layer between a company and its AI agents. It is planned for teams of 20 to 2,000 already using two or more AI assistants.
Planned: remote MCP, a consented browser extension, REST API, and web console.
MVP planned for month 4 after kickoff.
Planned · one memory graph per workspace
ClaudeChatGPTGrok
Continuum
Governed shared memoryFactsPreferencesDecisionsProject state
For teams already using two or more AI assistantsScroll to explore
The problem
Context stops at the edge of each assistant.
Repeated re-briefing keeps context from compounding. Colleagues' agents stay invisible to one another, native memory is personal, model switches lose memory, and exports are only snapshots. Per-vendor silos also mean copied context documents and custom instructions maintained separately in each tool. Admins lack visibility and control to see, approve, retain, or delete memories—a compliance concern under GDPR and sector rules.
A colleague’s agent knows something yours does not
What one person’s assistant learns is not available to the rest of the team.
Native memory is personal
Vendor memory is tied to a person and product, not shared decisions or team scopes.
Switching tools leaves memory behind
Moving between models can lose context; exports are one-time snapshots, not a live link.
Admins lack visibility and control
Without a way to see, approve, retain, or delete memories, teams face a compliance concern under GDPR and sector rules.
Instructions live in too many places
Per-vendor memory silos sit alongside copied context documents and custom instructions maintained separately in each tool.
How it works
One governed memory graph for the workspace.
The plan captures business facts, preferences, decisions, and project state, consolidates them into one graph per workspace, then gives each connected agent the relevant context it can access.
See what is planned by releaseFour planned MCP tools
A small, consistent set of actions
remember
Store a typed memory candidate with its subject, scope, and source.
recall
Return ranked memories the requesting agent is allowed to read.
context_pack
Build a brief for a person, project, or client within a token budget.
forget
Retract a memory and propagate its deletion when a person or admin asks.
Planned memory lifecycle
Every memory has a source, scope, and way to change.
MVP capture and recall are sequenced before nightly consolidation and decay in v1. The stages below are planned capabilities, not a description of a launched product.
- 01MVP
Capture with consent
An agent, connector, or enabled browser site submits content only with the user’s consent.
- 02MVP
Extract typed candidates
A model turns submitted text into typed candidates; the original episode is kept as a pointer.
- 03MVP
Check and link
Candidates are checked for duplicates, updates, and contradictions, then linked to relevant people, companies, or projects. PII rules are part of verification; detection and redaction are planned for v1.
- 04MVP
Store with provenance
The memory keeps its source, author, time, scope, and history; each write is audited.
- 05v1 planned
Consolidate nightly
Duplicates merge, newer dated evidence supersedes older versions, and unresolved conflicts go to review.
- 06v1 planned
Decay by workspace policy
Unused, low-confidence memories can be archived after a workspace-set period; pinned memories do not decay.
- 07Planned
Learn from use, then recall
People and agents can rate a recall to adjust confidence. Recall ranks relevance, recency, confidence, and scope within the caller’s token budget.
Integration maturity
Each agent has a planned, vendor-specific path.
Listings and connectors are not presented as launched.
Claude
The planned path is remote MCP with OAuth. A Claude Connectors Directory submission is planned for v1; the extension is a web-chat fallback.
ChatGPT
A planned ChatGPT App uses Apps SDK and MCP tools. Workspace admins must enable apps; listing follows developer-mode submission. The extension is a chatgpt.com fallback.
Grok
The planned path uses the same MCP server and a setup guide. The connector flow is still maturing; the extension is a grok.com fallback.
In-house agents
The plan uses REST for in-house agents. Python and TypeScript SDKs plus OpenAI Agents SDK, Claude Agent SDK, and LangGraph adapters are planned for v1.
Gemini and Copilot
These integration paths are deferred to the month 12 v2 plan.
Continuum is a working name. Product capabilities and release timing are plans, not shipped features.
Roadmap
A staged product plan.
The MVP is planned for month 4 after kickoff; paid Team plans are planned from month 6. v1 follows at month 8 and v2 at month 12. These are roadmap targets, not achieved milestones.
Planned · month 4 after kickoff
MVP
Prove that a memory created in one agent is usefully recalled in another within a day of setup.
Capture and memory
- remember, recall, forget, and context_pack MCP tools; remote MCP and REST API.
- Consented browser extension for chatgpt.com, claude.ai, and grok.com, with context-pack injection.
- Typed extraction, deduplication, hybrid semantic and keyword recall, and personal, team, and workspace scopes.
Console and governance
- Timeline, search, edit, pin, and forget; connected-agent visibility shows what each agent can see.
- Owner, admin, and member roles; encryption in transit and at rest; workspace tenant isolation.
- Immutable read/write audit, delete on request, and JSON/Markdown export.
Billing
- Free and Pro self-service billing; paid Team plans are planned from month 6.
Planned · month 8
v1
Connectors and memory
- Slack, Gmail, Notion, and Google Drive connectors, plus meeting-note import.
- Nightly consolidation, versioned conflict resolution, decay, and confidence scores.
Delivery and teams
- Python and TypeScript SDKs; OpenAI Agents SDK, Claude Agent SDK, and LangGraph adapters.
- ChatGPT App listing after developer-mode submission; planned Claude Connectors Directory submission.
- Team spaces, a shared-memory approval queue, sharing rules, and Team seat management.
Governance
- SAML SSO, SCIM, retention policies, PII redaction, and per-tenant encryption keys.
Planned · month 12
v2
Capture and memory
- CRM and ticketing connectors, webhooks, and bulk context-document import.
- Memory-quality analytics, per-agent recall tuning, and a weekly change digest.
Agent and admin tools
- Agent-to-agent handoff packets, offline export bundles, and Gemini/Copilot paths.
- Admin analytics, policy templates, and per-agent usage reports.
Enterprise
- EU/US data residency, VPC, DLP integrations, and a planned SOC 2 Type II report.
- Enterprise contracts and usage-based API billing.
Plan-wide privacy and safety commitments
Governance is part of the product plan.
- Customer memories are not used to train models.
- Model providers are under data-processing terms, with zero retention where offered.
- Memories are reference data, never instructions.
- Writes from webpages, emails, and files stay quarantined until a person or trusted agent confirms them.
- The browser extension captures only the user’s own conversations on enabled sites and shows a visible active indicator.
- Every person can inspect what the system knows about them and remove any item.
These are commitments in the plan; future controls and assurance are not represented as currently available or certified.
Who it's for
For teams that explain the same context every day.
First customers are teams already paying for at least two AI assistants and repeatedly re-explaining context. The initial focus is teams of 20 to 2,000; regulated enterprises with 2,000+ staff are a later audience, not the initial core.
Professional services
20–500 staff
Agencies, consultancies, boutique law, and accounting firms, where client context is the work and repeated briefs cost billable time.
AI-forward product and engineering
50–2,000 staff
Software teams building internal agents across models and needing shared memory with permissions.
Operations, RevOps, and chief of staff
200–2,000 staff
Mid-market teams trying to find decisions and preferences buried in chats.
Later enterprise audience · 2,000+ staff
Regulated enterprises
Finance, health, and public-sector teams may have later governance needs. They are not the initial core segment.
Who evaluates and uses it
Different roles need different proof.
Champion
Ops leads and account directors want fewer repeated briefs across agents.
Builder
AI and platform engineers want one MCP server, SDKs, scopes, and audit.
Buyer
Heads of AI, CIOs, and COOs look for neutrality, visibility, and exportable data.
Gatekeeper
Security, legal, and data-protection teams need clear data-flow information and review DPA, subprocessor, audit, retention, redaction, and assurance commitments by stage.
End user
Knowledge workers need to see what is remembered about them and remove any item.
Early access
Tell us about your team.
Share your work email, company, and team size to save your interest in Continuum. No automatic email or follow-up will be sent.
For teams working across two or more AI assistants.