Back to Articles

Essay

MCP from the wire up: give an assistant project memory

Follow a decision lookup from tool discovery to a usable answer, then see what changes when the same assistant can write.

November 22, 20263 min readMCPagent skills

A developer asks, “Why does this service use a queue?” The answer lives in a decision note from six months ago. An assistant with repository access could inspect the current implementation and still miss the reason it exists.

My OneCue skill teaches a coding agent when a previous decision is worth surfacing. Imagine exposing the underlying lookup to a second editor. The Model Context Protocol (MCP) gives that editor a way to discover and invoke a tool. The lookup below is a proposed adapter for this example, separate from the published skill.

Where the lookup runs

Editor hostConversation, user controls and model
MCP clientConnection to the memory server
Memory serverSchema, access checks and lookup
Decision notesRepository-scoped Markdown
The editor owns the interaction. The server decides which repository records this connection can read.

Put the matching logic beside the data. A result can carry the decision, its rationale, and the path to the original record. The host can then offer a source link. A model that gets only “use a queue” has no way to tell whether that advice still applies.

Discover, then call

This trace uses the 2025-11-25 protocol revision. Initialization negotiates the version and capabilities. After the server responds and the client sends its initialized notification, a client can list the tools advertised by a server with the tools capability. A paginated listing can require more than one request. Lifecycle · Tools

Only the last exchange searches the notes. Discovery tells the editor how to construct that request.Editor client to Memory server: initialize. Memory server to Editor client: version + capabilities. Editor client to Memory server: notifications/initialized. Editor client to Memory server: tools/list. Memory server to Editor client: names + schemas. Editor client to Memory server: tools/call. Memory server to Editor client: decision + source.Editor clientMemory serverinitializeversion + capabilitiesnotifications/initializedtools/listnames + schemastools/calldecision + source
Only the last exchange searches the notes. Discovery tells the editor how to construct that request.

Suppose the listing describes a tool named find_decision with a required string argument named query. Here is one complete JSON-RPC request and its response. The tool name and record are illustrative.

{"jsonrpc":"2.0","id":7,"method":"tools/call","params":{"name":"find_decision","arguments":{"query":"queue before sending notifications"}}}
{"jsonrpc":"2.0","id":7,"result":{"content":[{"type":"text","text":"Decision 14: queue notifications so a provider outage cannot block checkout. Source: decisions/014.md"}]}}

The host matches the response to request 7. It can show the rationale and let the developer open the note before changing the code. If no note matches, a successful empty lookup is useful information. Returning a vaguely related decision would quietly turn an evidence lookup into a recommendation.

What should a retry do?

ReadFind a note
ProposePreview an edit
WritePersist once
Adding a write tool introduces an application decision between finding evidence and changing it.

Consider a second tool, save_decision. The user approves a note, the server writes it, and the connection drops before the response arrives. Issuing another JSON-RPC request with a fresh ID can create two notes. Request IDs correlate responses. If duplicate writes matter, give the operation a durable application key and have the server check that key before inserting.

Test three outcomes separately: an absent note, a tool execution failure reported in a result with isError, and a JSON-RPC error for an unknown method. They call for different UI: continue without a cue, explain why the lookup failed, or fix the client integration. For work that another agent owns over time, follow the A2A task walkthrough.