A local-first MCP memory server for your agents.

The Model Context Protocol (MCP) is an open standard that lets AI agents call external tools through a uniform interface. An MCP memory server is one such tool provider: instead of returning weather or search results, it stores and retrieves durable context — decisions, files, past sessions — so the agent can remember across conversations.

Most agents have no long-term memory: Claude Code, Codex and Cursor each write history to their own files and never read the others'. An MCP memory server sits beside them and answers questions like “what did we decide about auth in this repo?” or “resume where the last session stopped” — for any client that speaks MCP, from one shared store.

get_project_context

Returns a condensed brief for a project: active decisions, recently touched files, repo conventions and the last errors encountered — scoped by repo, branch and Git worktree.

search_memory

Full-text and semantic search across every indexed agent session and saved note, each result carrying its provenance (which agent, which session).

save_memory

Persists a structuring decision — an architecture choice, an API contract, a validated workaround — so the next session, in any agent, can find it again.

ask_memory

Answers a natural-language question grounded in your indexed history, with citations back to the source sessions rather than a hallucinated summary.

prepare_handoff

Generates a context capsule — goal, decisions, Git state, tests, next verifiable steps — to hand a task from one agent to the next without pasting chat history.

1. Run the server locally

Download the signed Tramya app (macOS, Windows, Linux) from tramya.com. It starts an MCP server on 127.0.0.1:4317 and indexes the local history in ~/.claude/projects/ and other agents' stores.

2. Register it in one click

From the dashboard, click “Install in my agents”. Tramya writes the mcpServers entry into each client's config — ~/.claude/settings.json for Claude Code, and the equivalent for Codex, Cursor and Cline — so no manual JSON editing is required.

3. Call the tools

Add one line to your global rules (e.g. CLAUDE.md): call get_project_context at the start of a task, save_memory before it ends. From then on the agent reads and writes memory over MCP with no further setup.

Many MCP memory servers are hosted: they store your context — including snippets of proprietary code and internal decisions — on a third-party service and often bill per token. Tramya is local-first: the SQLite index lives in ~/.tramya/, search runs offline, and nothing leaves the machine unless you explicitly turn on Cloud sync. When you do, the vault is encrypted device-side (AES-256-GCM, key derived from a passphrase with scrypt) before anything is uploaded — useful for sharing project context across a team without sharing raw prompts.

The trade-off is simple: a hosted server is zero-install but exposes your history; a local-first server keeps the data on your disk and works in a train or a plane. If you want a broader survey of the options, see the guide to MCP memory servers.