create_agentic_sessionLABSCO SUMMARY
Three mechanisms, all aimed at the same waste.
Deduplication (store once, reference by key), deltas (send only what changed since a version), and lazy loading (expand a section only when a worker asks) are separate ideas that attack one problem — workers re-receiving context they already have. The dependency tools are the quieter half: declaring an output before it exists lets a worker wait on it instead of polling or duplicating the work.
What it is
An MCP server that gives an agentic team a shared memory instead of a copy of the context per worker. Context is stored once and referenced by key, workers claim units of work, discoveries are appended and shared, and updates are sent as deltas from the version a worker last saw.
What you get
- A session created for a coordinator and its workers, with session state readable and updatable —
create_agentic_session,get_session_info,update_session_status - Compressed context for a worker — a summary plus a reference key and expansion hints instead of the full object —
get_worker_context, withexpand_context_sectionto pull a section on demand - Only the changes since a version a worker already has —
get_context_deltawithsince_version - Work units published, claimed with an estimated duration, and progressed —
publish_work_units,claim_work_unit,update_work_status - Findings shared as they happen, addressed at the workers they affect —
add_discovery,get_discoveries_since - Dependency resolution between workers —
declare_outputs,await_dependency,publish_output
Requirements
No account and no key — it runs locally. Node.js 18+ and npm or yarn. Clone the repository, npm install, npm run build, then point Claude Desktop at dist/server.js with node; the repository ships claude-desktop-config.example.json to copy. The package is shared-memory-mcp 1.0.0, MIT licensed. npm run test-workflow exercises the full agentic flow.
Setup effort
Build from source — git clone https://github.com/evalops/shared-memory-mcp.git
USE CASES
01
Briefing four workers without paying for four full contexts
Store the context once and hand each worker a summary plus a reference key, expanding only the sections they actually open.
02
Keeping a long-running worker current
Ask for changes since the version it last saw rather than re-sending the whole state.
03
Splitting work without two agents grabbing the same task
Publish the units with their dependencies, and let each worker claim one with an estimated duration.
04
Telling the rest of the team what you just found
Append the discovery and name the workers it affects, so it reaches them without a broadcast.
05
Waiting on something another worker has not produced yet
Declare the output up front and await it, instead of polling or rebuilding it in parallel.
14 tools
Listed in project docs. Your client may show a different list.
get_session_infoupdate_session_statusget_worker_contextexpand_context_sectionget_context_deltapublish_work_unitsclaim_work_unitupdate_work_statusadd_discoveryget_discoveries_sincedeclare_outputsawait_dependencypublish_outputBEFORE YOU INSTALL
What to know first
Tools you get
14 tools named in the project's documentation.
Requires
Node.js 18+
Where it runs
The documented client runs the server on your machine.
INSTALLATION
Connect your client
You clone and build this one yourself.
Server setup
TERMINAL
git clone https://github.com/evalops/shared-memory-mcp.git
cd shared-memory-mcp
npm install
npm run buildMCP Clients
Claude Desktop
CLAUDE DESKTOP · JSON
{
"mcpServers": {
"shared-memory": {
"command": "node",
"args": [
"<YOUR_PATH_TO_SHARED_MEMORY_MCP>/dist/server.js"
]
}
}
}
