Memwright MCP Server

Cut your agent's token burn 21×. Two API calls. Full-context replay re-reads the whole conversation every turn - input tokens that grow O(n²) and a bill that compounds with every session. Attestor retrieves only what's needed: flat ~200 tokens per call, 21× fewer input tokens by turn 100, 100% recall - measured across six models, open and closed. It runs locally over stdio via the published package.

People connecting memory tools to Claude, Cursor, VS Code, or another MCP client. The project is written in Python.

VERIFIED ACTIVE

LAST COMMIT 2026-05-30 · ★ 14 · #106 OF 259 MAINTAINED MEMORY · VERIFIED 2026-08-25

MIT · Python servers · how we verify → /methodology

01 · Install Memwright

Claude Code

claude mcp add bolnet-agent-memory -- uvx memwright

Claude Desktop / Cursor / VS Code - add to config

{
  "mcpServers": {
    "bolnet-agent-memory": {
      "command": "uvx",
      "args": [
        "memwright"
      ]
    }
  }
}

Same JSON for Cursor. For VS Code, rename the top-level key from `mcpServers` to `servers`.

Using another client? Same JSON, different key

Claude Desktop · mcpServers

Cursor · mcpServers

VS Code · servers

Windsurf · mcpServers

Zed · context_servers

Cline · mcpServers

Roo Code · mcpServers

Continue · mcpServers

LibreChat · mcpServers

Gemini CLI · mcpServers

Codex CLI · mcp_servers

Full setup guides: every client.

02 · Evidence

Security posture

What to check before giving this server access to your agent - from the registry, GitHub, and our own probes. We don't score safety; we show what's verifiable.

runs as local process (stdio) - runs on your machine with your user's permissions

license MIT - declared in the repository

pypi package memwright - check the name against the project README before installing (PyPI has no namespace ownership)

registry namespace io.github.bolnet is GitHub-verified and matches the repo owner

03 · What Memwright can do

Prose above is summarized from the project's README and registry record - no invented capabilities.

Latest releases

v4.1.11 · 2026-05-26

Adds attestor recall --show-tokens: prints the packed token count of the recall payload so per-call recall size is visible (flat regardless of conversation length). Ships alongside the 4.1.10 uncapped budget.

v4.1.10 · 2026-05-26

Reverses the 2048 cap (4.1.9): recall budget set to 1,000,000 (effectively uncapped; still bounded by vector_top_k) so we observe real token usage before tuning. Applied to MCP default, CLI --budget, init config…

v4.1.8 · 2026-05-26

Every Docker container, volume, network, and the compose project is now named attestor_… - so docker ps -a | grep attestor, docker volume ls | grep attestor, and docker network ls | grep attestor list everything…

04 · Who maintains Memwright

memwright is maintained by bolnet. We track 2 MCP servers from bolnet - 2 actively maintained, 28 combined GitHub stars, oldest repo from Mar 2026. Full record: all servers from bolnet.

  1. Attestor Self-hosted memory for agent teams. Bi-temporal replay, deterministic retrieval, audit log. ★ 14

05 · Facts

category
memory - ranked #106 of 259 actively-maintained memory servers as of 2026-08-25.
registry
io.github.bolnet/memwright (active, first published 2026-03-07 · 2 versions)
packages
pypi:memwright

06 · Memwright FAQ

What is Memwright?

Cut your agent's token burn 21×. Two API calls. Full-context replay re-reads the whole conversation every turn - input tokens that grow O(n²) and a bill that compounds with every session. Attestor retrieves only what's needed: flat ~200 tokens per call, 21× fewer input tokens by turn 100, 100% recall - measured across six models, open and closed. It runs locally over stdio via the published package.

Is Memwright still maintained?

Yes - as of 2026-08-25, its last commit was 2026-05-30. We re-verify nightly.

How do I install Memwright?

Run `uvx memwright`. You can also paste the ready-made client config above.

Does Memwright run locally?

Yes - it's a stdio server: it runs on your machine (via uvx) with your user's permissions. Your data stays local unless the server itself calls external APIs.

07 · Alternatives to Memwright

Pairs well with

Servers that cover what Memwright doesn't - only shown when the pairing reason fits the companion.

  1. Server Filesystem → filesystem companion filesystem · ★ 38,704
  2. X402 List MCP → filesystem companion filesystem · ★ 58