Central service
The authoritative service holds the tools and their semantics, authorisation decisions, project scope and persistence. It is the only process that touches the database.
How it works
People, agents and models sharing one durable workspace.
SHAMPOO keeps memory, tasks, context and decisions in one place every participant can reach, with boundaries, history and human approval built in.
The working loop
Four steps, repeated for every piece of work.
-
Step 1
Connect through your own boundary
Each assistant connects to the central service through a thin local adapter. The connection selects one profile (one independently authorised memory boundary) and the session stays bound to it.
-
Step 2
Work against shared state
Tasks, notes, entities, observations and relations live outside any single chat. A fresh agent session sees what is ready, what is in progress and what was already decided, instead of starting blank.
-
Step 3
Bring humans in where it matters
Ambiguous context can be queued for a person, whose answer becomes part of the durable record. Inbound material can require review before it becomes agreed state. Nothing autonomous is hidden from the operator.
-
Step 4
Agree, record and resume
Positions are debated in the open, agreements are recorded with named participants, and work streams are claimed, completed and handed over. Later sessions resume from the record, not from rumour.
Separate strands in, one durable structure out
Who works in SHAMPOO, and what their work becomes.
- Human
- Agent
- Model
- Automation
SHAMPOO
- Memory
- Tasks
- Knowledge
- Sessions
- Decisions
- Workflows
The three parts
One service, one adapter, one mechanism.
Workstation adapter
A deliberately thin local process presents an MCP interface to assistants, owns the bootstrap secret, renews short-lived sessions and forwards everything sealed. It holds no authority of its own: remove the central service and the adapter is useless.
Delivery mechanism
A separate mechanism materialises recurring timers and drains delivery jobs with retry and backoff. It transports work; it never decides direction, agreement or approval.
Nothing is silently lost
State moves; history stays.
Tasks move, they are not deleted
Work travels not started → in progress → done, with archived and cancelled as terminal states. Every transition preserves history.
Knowledge retires, it is not erased
Entities, observations and relations move through lifecycle states rather than being physically deleted, and retired knowledge can revive with its original evidence intact.
Sessions expire and renew
Session credentials are short-lived and renew automatically; revocation takes effect on the very next operation.