🧭 Mindtrail
Leave a trail your AI agents can follow.
Persistent, local-first memory for coding agents over MCP. Tell your agent something once, and every future session remembers it.
Усильте агента
Собственная команда издателя. Не эндпоинт PlugMCP.
Host a copy with our keys and logs.
uvx mindtrail==0.1.1У Claude Code нет ссылки для установки локальной команды. Скопируйте команду ниже и выполните её в терминале.
claude mcp add io.github.rohitdeshmukh-1-mindtrail -- uvx mindtrail==0.1.1Выполните в терминале.
{ "mcpServers": { "io.github.rohitdeshmukh-1-mindtrail": { "command": "uvx", "args": [ "mindtrail==0.1.1" ] } }}claude_desktop_config.json
Удалённый URL или команда на вашем компьютере? Чем различаются два способа установки
Запустить на PlugMCP
A new server you create. It is not this listing's official endpoint. Use the publisher install when you want their URL and their credentials.
ЗапускОт издателя
Install · Day to day · How it works · Benchmarks · Docs
Every new agent session starts from zero. You explain your conventions again, re-state your preferences, and re-describe decisions you made last week. Mindtrail gives your agents a shared, long-term memory through the Model Context Protocol, so what one session learns, every later session (in any tool) can recall.
uv there is
nothing to install first.We tested it with real Claude Code sessions. In one session the user mentions a project fact in passing ("FYI, this project uses pnpm"). A fresh session later gets a task that depends on it ("how do I add lodash?"). The agent is never told to use Mindtrail.
Without memory, the agent answers from habit every time: npm install lodash, a commit
message in the wrong format, port 5432 instead of your 5433. With Mindtrail it checks first
and gets your project's answer. This is a small test (12 facts run three times, one model,
an empty repository), not a measure of task success on large codebases.
Method, raw results and how to reproduce.
Mindtrail needs Python 3.11+. The fastest route is uv: uvx
downloads and runs Mindtrail on demand, so you only add it to your client. The first time the
server starts, it downloads the embedding model (~210 MB) in the background.
claude mcp add mindtrail --scope user -- uvx --from "mindtrail[semantic]" mindtrail serve
--scope user makes Mindtrail available in every project. Check it with claude mcp list.
Add this to the client's MCP config (in Claude Desktop: Settings → Developer → Edit Config):
{
"mcpServers": {
"mindtrail": {
"command": "uvx",
"args": ["--from", "mindtrail[semantic]", "mindtrail", "serve"]
}
}
}
# ~/.codex/config.toml
[mcp_servers.mindtrail]
command = "uvx"
args = ["--from", "mindtrail[semantic]", "mindtrail", "serve"]
Install the CLI once with pipx (or pip), then point your client at mindtrail serve:
pipx install "mindtrail[semantic]"
mindtrail init # downloads the model now and prints client configs
claude mcp add mindtrail --scope user -- mindtrail serve # Claude Code
Want the smallest install? Drop [semantic]: Mindtrail then needs no model download and
matches on words instead of meaning (see Better semantic recall).
More clients and troubleshooting are in docs/integrations.md.
Session 1 › FYI, this project uses conventional commits and squash merges.
Session 2 › Write a commit message for these changes.
→ the agent recalls the convention and writes "feat(api): add pagination to /orders"
uvx picks up new releases on its own (uvx --refresh ... forces a check). With pipx, run
pipx upgrade mindtrail. Your memories live in ~/.mindtrail and are kept across upgrades.
See the changelog for what changed.
You don't need special commands. Work as usual and the agent decides what to keep:
release branch", "I prefer pytest over
unittest", "the flaky test was a timezone bug, fixed by pinning TZ=UTC". The agent stores
facts like these on its own. Say "remember that…" when you want to be sure.mindtrail list and mindtrail recall "<question>".mindtrail forget <id>. Deletes
are permanent.Good things to store: conventions, commands, ownership, where config lives, decisions and why they were made, root causes of tricky bugs. Don't bother with what the code or git history already says. Secrets are refused automatically.
Your agent gets three tools, and Mindtrail's server instructions tell it when to use them:
Set MINDTRAIL_TOOLS=full for four more: search_memory (filters by space, type and validity),
get_context (a prompt-ready block within a token budget), update_memory (edits with
version history) and get_memory.
Behind the tools is a small, well-tested engine: duplicate merging, supersession, validity windows, version history and hybrid ranking. See docs/architecture.md.
mindtrail recall "how do we deploy?" # search project + personal memory
mindtrail remember "Staging is at staging.example.com" --scope project
mindtrail list # newest first
mindtrail forget <id> # permanent delete
mindtrail export -o memories.jsonl # everything you've stored, as JSON Lines
mindtrail doctor # diagnose the install
The default embedder matches on words and word fragments. For recall by meaning ("how do we deploy?" → "deploys go through GitHub Actions"), install the local neural model. It runs on CPU and needs no API key:
pipx install "mindtrail[semantic]" # or use the uvx config above
mindtrail init # downloads the model once (~210 MB) and re-indexes existing memories
For a smaller download (67 MB, somewhat lower recall), set
MINDTRAIL_EMBEDDING_MODEL=BAAI/bge-small-en-v1.5. Memories stored under another model are
re-embedded automatically the next time the server starts.
The agent test is in Does it actually help?. On two held-out sets of developer-memory questions (82 in total) that were never used for tuning (methodology):
This is a small, synthetic retrieval benchmark written by us. Run mindtrail bench to
reproduce it, or add your own dataset.
from mindtrail import MemoryService
memory = MemoryService.from_config()
memory.remember("The API uses FastAPI and PostgreSQL", space_id="project:shop")
print(memory.get_context("add a new endpoint", space_ids=["project:shop"]).text)
Everything stays in ~/.mindtrail/mindtrail.db on your machine. Mindtrail refuses writes that
look like API keys, tokens or private keys. forget overwrites deleted data on disk, and
recalled memories are marked as untrusted reference data so agents don't follow instructions stored inside them.
See SECURITY.md to report issues.
Issues and PRs are welcome. Read CONTRIBUTING.md to get set up; the whole suite runs in about 15 seconds. If Mindtrail saves you from re-explaining your codebase, a ⭐ helps others find it.
Persistent, local-first memory for AI coding agents. Remember, recall and forget across sessions.
{ "mcpServers": { "io.github.rohitdeshmukh-1-mindtrail": { "command": "uvx", "args": [ "mindtrail==0.1.1" ] } }}Инструменты
Списка инструментов пока нет. У удалённых серверов он появляется после первой проверки состояния.
FAQ
Скопируйте на этой странице конфигурацию stdio для Cursor, Claude Desktop, Claude Code или VS Code. Она использует установку издателя, а не прокси PlugMCP.
В этой записи не указано обязательных переменных окружения.
Mindtrail работает локально через stdio.
Последняя проверка: 2026-11-09. Persistent, local-first memory for AI coding agents. Remember, recall and forget across sessions.