Evidence that the need is real
| Signal | What happened |
|---|---|
| Vendors | Anthropic launched memory for Claude Managed Agents (public beta, 23 April 2026, with audit trails); OpenAI and Google Cloud followed within weeks; Microsoft added Copilot Memory and organisation-level "Work IQ". |
| Analysts | PitchBook's 2026 AI Outlook is reported to call shared memory the defining competitive battleground; Futurum asks whether shared memory becomes a standard tender (RFP) requirement by 2027. |
| Early results | Rakuten reports its Claude agents make 97 % fewer first-pass errors at 27 % lower cost and 34 % lower latency because they remember earlier mistakes (vendor-reported case). |
| New entrants | Writer launched "Enterprise Brain", a governed context layer for agents (early access, 9 September 2026); Tabnine offers a permission-aware Context Engine for code. |
Who it is for
Strong fit
- Mid-size and large companies where many people and AI agents work on shared projects
- Consultancies and suppliers with many clients that must be kept apart
- Regulated or export-controlled industries — shipbuilding, automotive, finance, health
- Companies using several AI tools (Claude, Copilot, ChatGPT, local models)
- Long programmes and staff turnover, where knowledge walks out of the door
Weak fit
- Very small teams where everyone already knows everything
- Companies that hardly use AI yet
- Companies fully committed to one vendor's suite, where built-in memory may be enough
- Mostly one-off work with little to reuse
What exists today: four groups
| Group | Examples | Good at | Missing for a company-wide memory |
|---|---|---|---|
| 1 · Developer memory kits | Mem0, Zep, Letta, Hindsight (used inside cMCP) | Storing and finding memories for one app or agent; MCP servers; self-hosting possible | No company governance — no scopes, approvals, client walls or lineage. Building blocks, not a company system. |
| 2 · Memory inside one AI vendor | Claude Managed Agents, Microsoft 365 Copilot (Work IQ), ChatGPT, Gemini | Easy to switch on, deeply integrated, often included in the price | Tied to one vendor — memory does not follow you to another model; mostly cloud only; little governed sharing up through teams and projects. |
| 3 · Company search | Glean, Atlassian Rovo / Teamwork Graph | Finding what is already written across documents, mail and tickets, respecting permissions | Searches existing content; does not learn new knowledge from AI work or move it between scopes. |
| 4 · Governed context layers | Writer Enterprise Brain, Tabnine Context Engine | Shared, governed context for many agents; Tabnine also on-premises / air-gapped | Writer centres on brand and marketing content inside its own platform; Tabnine on code only; auditable controls for regulated sectors still to be proven. |
Side by side
| Capability | Memory kits | Vendor memory | Company search | Context layers | cMCP |
|---|---|---|---|---|---|
| Works with any AI model | yes | no | partly | partly | yes |
| Learns new facts from AI work | yes | yes | no | yes | yes |
| Scopes: person → team → project → company, with approval | no | partly | no | partly | yes |
| Client walls (many clients kept apart) | no | no | partly | partly | yes |
| Runs on own servers / without internet | yes | no | partly | partly | yes |
| Audit trail of every read, write and approval | partly | partly | partly | yes | yes |
| Mature, proven at large scale | partly | yes | yes | no | not yet |
Group-level view from public information; individual products differ and change quickly. "Partly" means some products in the group offer it, or only in a limited form.
Where cMCP is different
- Any model through open standards (MCP, REST) — no lock-in, which analysts list as a main risk of vendor-bundled memory.
- Governed flow across six scopes — knowledge moves up only with a lead's approval and with names removed, flows down automatically, and is combined across teams.
- Client walls and multi-organisation accounts — built for consultancies and suppliers serving several large clients.
- Own servers, own model — runs on premises or without internet, with the language model in-house.
- Audit and approvals built in — four-eyes for connections, approval cover, every read and write recorded.
Honest risks
- Free bundled memory. Many companies will accept "good enough" memory inside Copilot or Claude. Our case is strongest for regulated companies using several models.
- Fast-moving market. Writer, Tabnine and the big vendors are heading here; a governed, model-independent product could appear within 12–18 months.
- Stage. cMCP Phase 1 (stages 1a–1t) is built and tested on one pilot server, but no company uses it in production yet; it moves to IT-Labs hardware next. Large scale must be proven by load tests (see the scalability fact sheet).
- Borrowed engine. The search engine underneath is open source (Hindsight, MIT); our value is the governance layer, which must stay clearly ahead.
Sources
- Writer's Enterprise Brain adds shared memory for agents (CMSWire)
- Will shared memory become the missing link for enterprise multi-agent AI? (Futurum)
- Built-in memory for Claude Managed Agents (Anthropic)
- Claude Managed Agents memory 2026 (Build Fast with AI)
- Best AI memory systems in 2026 (EverMind)
- Best AI agent memory frameworks 2026 (Atlan)
- Best AI agent memory providers 2026 (Developers Digest)
- Best MCP memory servers 2026 (Hjarni)
- Microsoft 365 Copilot Memory enterprise guide
- What's new in Microsoft Copilot for 2026 (Wavenet)
- Glean vs Microsoft Copilot (AI Agent Rank)
- Enterprise AI memory: the 2026 guide (Sentra)