MCPmemory · skills · handoffs

Personal memory is solved. This is the shared one.

Give Claude, Codex and any MCP client — and the people working beside them — shared knowledge, reusable skills, and a reliable way to hand work to each other.

Shared across every MCP client Hosted or cloud models Update one skill for every agent On the official MCP Registry: com.levirge/brain

levirge brain ● your workspace · MCP

Levirge Brain — the entity browser with demo data: an Ask bar, Skills / Entities / Docs / Metrics tabs, a tracked-entity sidebar, and a cited per-entity summary

MCP tools from one server — knowledge, skills, handoffs

8 → 1

eight search fragments become one cited summary — see the example

third-party model calls in hosted mode

The problem

Agent memory stops at four walls

Built-in memory stays with the product that collected it. Brain sits outside all four walls, so every tool, model, teammate and agent reads one knowledge base.

Tool

Codex doesn't know what you settled in Claude Code, so you brief it again.

Every MCP client reads one endpoint.

Model

Switching provider means rebuilding your context from zero.

Memory lives outside the model.

Teammate

A colleague's agent repeats an investigation yours finished last week.

One shared Vault per project.

Agent

A new session can't tell what the last one was doing or still owes.

Handoffs, sessions and todos.

Anthropic's own docs: “Each Claude Code session begins with a fresh context window.” (source, checked 29 Oct 2026). What people do instead, and where it stops →

Handoffs

Pick up where another agent left off

Agents hand each other work through a durable inbox, with the context attached. A new session asks what it was doing, what's still owed, and what someone left it.

See the full loop →

handoff · inbox durable

eveval@brain→ brain@levirge

3 defects from the scifact benchmark run

waited in the inboxdelivered on connect
brbrain@levirgereply

Fixed in PRs #133 and #135, both merged.
eveval@brainreply

Re-ran the benchmark against the fixes.

verified

A real loop from 15 Aug 2026: a benchmark agent hands three defects to the agent that maintains Brain, then checks the fixes.

Curation

One cited summary, not eight fragments

Brain groups what agents capture by entity and keeps a short summary of each. Every statement links to its sources, and the summary rebuilds when a source changes.

See the before and after →

Captured

session-capture.mddeploy-timeline.txtadr-0021-api-tokens.mdmcp-connectivity-drops.mdlistener-log.mdhandoff-thread.md

distilled

What the agent reads

entity · search-mcp

Redeploys kill in-memory MCP sessions and hour-long JWTs — clients reconnect with static per-user API tokens instead. Decided after the August connectivity incidents.

adr-0021-api-tokens.mdmcp-connectivity-drops.md+2

A real summary, distilled from our own August incident notes.

Teams

One Vault per project, shared by the team

Several people, each running their own agents, read and write the same Vault. Every document records whose agent captured it. Tokens can be limited to chosen Vaults and made read-only.

How access works →

vault · payments 3 people · 2 tokens

People

prPriyaNoneViewerMember

saSamNoneViewerMember

alAlexNoneViewerMember

Tokens

⌁ci-deployread-only

⌁release-botread · write

runbooks/deploy.md · captured by Sam's agent

Demo data. Access is None, Viewer or Member per person; a token can be limited to chosen Vaults and made read-only.

Obsidian

Your team curates in Obsidian

The Obsidian plugin two-way syncs a local vault with Brain. People edit in the notes app they already use; agents recall the same knowledge everywhere else.

Install the plugin →

obsidian · my-vault ● brain

decisions

pricing.md

pricing (brain conflict).md

runbooks

deploy.md ↑ pushed

rollback.md

rollback (brain conflict).md

entities

search-mcp.md ↓ pulled

Brain sync: 4 pulled, 1 pushed, 2 conflicts

Demo data, the plugin's real file names. A conflict never overwrites your note; Brain's copy is parked beside it.

Where it runs

Private by default, fast when you need it

Hosted mode runs every model call on LLMs Levirge hosts, with no third-party provider in the path. Switch to cloud LLMs for speed at runtime, with no redeploy.

See plans →

settings · model lane

Private

Processed on our own infrastructure. Content never leaves the network.

Fast

Processed by a hosted cloud service. Faster, but content is sent to a third-party API.

In Settings the hosted lane is called Private and the cloud lane Fast. Plans from $20/mo; $60/mo puts your corpus on the private lane.

Install

Pick your agent

Access is per workspace — sign up, then connect your clients.

claude plugin marketplace add levirge/brain
claude plugin install brain@levirge-brain

Installs the /brain:* commands and connects the MCP server. Restart Claude Code and sign in on first use. Token route for CI →

for agents ● MCP · Streamable HTTP

discovery: levirge.com/llms.txt
docs: levirge.com/brain.md
endpoint: https://brain.levirge.com/mcp
definition: levirge.com/brain/mcp.json
auth: OAuth sign-in · bearer token as fallback
verify: call kb_overview — read-only
tools: 20 — brain.md for agents, docs for humans

Next step

Give your whole fleet one memory

Shared knowledge, skills and handoffs for every agent and person on the team.