baodq06/memocall

An MCP server that lets Claude Code **recall the context of your past conversations — from any project — on demand.**
Claude Code can *resume* a session, but getting *another* conversation's context into the one you're in is a chore — copy-pasting exports or digging through raw transcript files by hand — and resume only sees the current directory anyway. **MemoCall** turns it into a single ask: *"what was I working on?"* or *"load the session where we set up X"*, and it pulls that transcript in as clean, compact Markdown — even from a totally different project.
It's **read-only**. Claude Code already records every session to ~/.claude/projects/ / .jsonl; this server just reads those files, strips the noise, collapses tool calls, and hands back a readable transcript. It never writes, hooks, or touches a live session.
What you get
Five tools, available in **every** session once installed. Claude picks the right one from the request:
| Tool | What it does |
|------|--------------|
| list_sessions | Your recent conversations across all projects, **grouped by directory**. "What was I working on?" |
| search_sessions | Find a past conversation by keyword (matches titles + first messages + paths). |
| load_session | Load one conversation as compact Markdown — by id or query. Optionally a turn window via turns. |
| session_outline | A cheap **map** of one conversation: the numbered list of your prompts. Great for huge sessions. |
| search_in_session | Return only the turns **within** one conversation that match a keyword. |
You don't call these directly — you talk normally and Claude reaches for them:
How the transcript is rendered
Raw transcripts are a verbose event log (one big file hit 11 MB). memocall reduces each turn to the essentials, Conductor-style:
Tool-call *outputs* are elided (the big token win, and a privacy win — see below). Abandoned/edited message branches are dropped so you get the conversation as it actually played out. A token budget keeps even an 11 MB session well under Claude Code's MCP output cap via middle-out truncation.
Navigating large sessions
A single response can't hold a 1,000-turn session, so for big ones you don't dump — you navigate:
1. session_outline → a numbered map of every prompt (tiny, fits any session).
2. search_in_session → jump straight to the turns about a topic, **or** load_session with turns: "300-340" to pull an exact window (turn numbers come from the outline).
So you never lose access to the middle of a huge conversation — load_session alone would middle-out-truncate it, but the outline + range/search tools let Claude reach any part on demand.
Install
Requires the Claude Code CLI (and Node 18+). One line — npx fetches and runs the published package:
--scope user makes it available in every session, in every directory. Restart Claude Code (or open a new session) and ask it to list your sessions. Verify with claude mcp list — you should see memocall: … ✔ Connected.