Engineering documents from the cMCP pilot build (October 2026). Statuses here are engineering states, not product claims: see cMCP for what is available. All documents

cMCP — Shared Cognitive Memory Layer

One memory for a company's people and their AI assistants: organised by organisation, team, project, client and person, permission-checked on every request, and kept up to date for everyone at once. This page is the front door to everything written about it.

Built: Build Phase 1, stages 1a–1t, tested on the pilot server Next: the move to IT-Labs hardware (M-112, on hold) Planning board v71 · 2026-10-11

Read and use

How to use it

The exact flow for the account admin, the user and the solo user: sign-in, daily steps, good habits and what to do when something goes wrong.

cmcp-how-to-use.html

Features and stack

Every part of the hardware and software, one tab each, with its flow, features and detailed usage notes.

cmcp-features-and-stack.html

How an AI app connects

The two ways in (a person's helper, or a bot with a sponsor), as flow charts, and the answer to "can a Claude app connect on its own?"

cmcp-ai-app-connection.html

What can be linked

Jira, GitHub, e-mail, databases and more: what is built, what is planned and what is not.

cmcp-what-can-be-linked.html

Open the application

The web console on the pilot server, until it moves to cmcp.it-labs.in. Sign in with your name, password and the one-time code from your authenticator app.

console.pilot.example

Plan and progress

Planning board v71

The single page with status, decisions M-01…M-112 (stage 1t built; the move to IT-Labs hardware planned; S-45 for 1u set aside), review, design, architecture, roadmap, risks, questions and changelog.

cmcp-planning-board-v76.html

Issues to discuss later

Build plan v1.2

Stages 1l–1t built; next: the move to IT-Labs hardware (console at cmcp.it-labs.in, the stack on a GPU server). The whole picture once more: what stages 1a–1k built, what the research (Q-35, Q-36) asks for, and the gap. Verdict: extend, don't rebuild. Shows how lanes, the work pool, collective and organizational memory fit the existing scopes, the new space registry, memory address, envelope and change feed, the cmcp helper with Claude Code hooks, and seven stages 1l–1r, each with an exit test. Eight decisions (N1–N8), proposal S-43.

cmcp-build-plan.html · plan v1.2 · 2026-10-11

Space and memory model (paper v1.1)

How cMCP maps the spaces people and their AI agents, bots and digital workers work in to the organization, so every memory answers which space, app, agent or bot, person, team, project and org. Covers four kinds of actor, space IDs, the space list sent to every app, session routing with a private work pool, collective and organizational memory, and building memory from mail, chats, documents and repos. Consolidates drafts v0.1–v0.3; v1.1 adds searchable indexes per category (people, agents and bots, places, spaces, topics, memory, knowledge, sources) and index cards loaded at session start like Claude's memory index; 24 decisions.

cmcp-space-memory-model.html · v1.1 · adopted and built as stages 1l–1t

User agent auto-mapping (Q-35)

How a person using any AI tool (Claude, ChatGPT / Codex, Gemini, Copilot, Cursor, Groq, local apps …) on many devices and projects gets their memory into the right place automatically. Research on about 40 tools; recommended: a small cmcp helper (setup, hooks, bridge, import), automatic project fingerprint, private spaces, ask-once linking, personal device keys, chat-app connectors and an allowed-AI-apps list per account. New on 2026-10-08: automatic saving and two-way sync for any agent or bot. Identity comes from the sign-in; a small envelope carries place and conversation; a cursor returns what changed; hooks and a bot library save at checkpoints, private first. Whole conversations can be kept only if company policy switches it on, under fixed default rules. New risk list: what the ADMIN and people are told before any switch that loosens a default goes on. Twenty-two decisions; the recommendation was adopted in the build plan (M-106) and built as stages 1l–1t.

cmcp-agent-auto-mapping.html · Q-35 · built in stages 1l–1t

Background papers

Architecture paper

Architecting a shared cognitive memory layer for enterprise AI infrastructure — matched to the software actually built, with the future-development parts kept.

enterprise-cognitive-memory.html · v1.3

Executive pitch

The management pitch for the shared memory layer, adjusted only where the build differs from the original plan.

enterprise-memory-pitch.html

Market and alternatives

Is this a real need, who is it for, what ready-made products exist in 2026 (memory kits, vendor memory, company search, context layers) and where cMCP fits.

cmcp-market-and-alternatives.html

Management pitch deck (PowerPoint)

24 slides presenting the idea to management, Big Idea first, with a spoken narration in the speaker notes for text-to-speech.

enterprise-memory-pitch.pptx · download

Measurements

Scalability fact sheet

Can the design serve a large shipbuilder or a Stellantis-sized car maker? What the core design already gives, what we build when needed, and indicative hardware.

cmcp-scalability-fact-sheet.html

LLM model results and timing

Which language models we tested for turning notes into facts: quality, speed, and how the time grows with the size of the text. Plus search and sync timings by data size.

cmcp-llm-model-results.html