The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Memocall listing page.
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/<project>/<id>.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.
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:
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.
A single response can't hold a 1,000-turn session, so for big ones you don't dump — you navigate:
session_outline → a numbered map of every prompt (tiny, fits any session).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.
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.
~/.claude/projects/ and nothing else.format: "full" includes brief tool inputs and may surface sensitive strings. Use it deliberately.cleanupPeriodDays (default 30).src/jsonl.ts so it's a one-file patch if it does.Layout:
src/jsonl.ts — all knowledge of the transcript format (helpers, ordering, classification).src/parser.ts — JSONL → compact Markdown (the core transform).src/locator.ts — session enumeration + cached metadata index.src/index.ts — the MCP server wiring the three tools.MIT