Desktop context over MCP

Pick up where your work left off.

MemoryRaven helps a compatible AI assistant recall recent work from your Windows PC. The agent collects text that supported apps make available to Windows accessibility tools, then sends it to our service for processing.

Designed for 64-bit Windows. The installer is not available yet.

How it works

The Windows agent reads supported on-screen text; the service prepares it for your assistant to request.

  1. Capture

    The Windows agent reads text and app details made available through Windows UI Automation by supported applications. What is exposed varies by app; this is not a screenshot or OCR capture service.

  2. Keep working offline

    While offline, a bounded queue on your PC holds events for a later upload. If its size or entry limit is reached, older queued events may be discarded. Activity logs and the queue are stored locally without encryption by MemoryRaven.

  3. Server-side pattern redaction

    The server applies pattern-based redaction to incoming event text before database storage. It cannot identify every sensitive value, and data is already present in local activity files before upload. Do not use MemoryRaven where this collection is not appropriate.

  4. Turned into something readable

    Rather than a firehose of window titles, each stretch of work becomes a short written summary. Your assistant gets the gist of recent activity in a few sentences, which is both easier to use and far cheaper than shipping raw activity around.

  5. Serve

    Ask what you were working on. Each assistant connection is limited to the computer you selected when authorizing it. The MCP tools are read-only; they do not control your PC.

The pipeline, end to end

For the technically minded: one agent on your machine, one append-only log on the server, one read-only tool surface for the model.

MemoryRaven architecture On your Windows PC, a UI Automation collector, the MemoryRaven agent and a bounded offline queue feed events over WSS to the MemoryRaven server. There the events are ingested and redacted, appended to a per-device hot buffer in PostgreSQL, sealed into an immutable segment with a monotonic sequence number, and then compressed by the configured model. Any MCP compatible MCP clients, including Claude or ChatGPT, can read the result through context_id and context_for_minutes. YOUR WINDOWS PC UI Automation collector 9 COM event types MemoryRaven agent tray + browser login Offline queue 50,000 entries / 50 MiB WSS MEMORYRAVEN SERVER Ingest, redact, append /ws/ingest Hot buffer PostgreSQL, per device, durable Sealed segment immutable, monotonic seq Model compression raw_list → compressed MCP CONSUMER MCP client context_id() context_for_minutes()
The core idea

Ask for new context without replaying everything

Context requests can use sequence numbers to distinguish previously returned activity from newer activity in the same session.

two calls, one session
context_id()        → seq 1182, 1183, 1184
                       + live hot buffer

  # compression rewrites 1184 in place,
  # the sequence number does not move

context_id()        → seq 1185
                       + live hot buffer

  # 1182-1184 already seen: not resent

Built to hand over as little as possible

You approve computers through account authorization. The Windows agent collects text exposed through UI Automation; the server applies pattern-based redaction before database storage, which may not catch all sensitive content.

Scoped per authorization

Each assistant authorization is scoped to the computer selected in your browser. You can approve more than one computer, so choose carefully when authorizing a connection.

Temporary raw diagnostic data

Raw event copies may remain for up to a day after successful processing; if processing fails, they may remain longer. MCP context can include recent activity as well as summaries. Filtering is pattern-based and cannot guarantee that all sensitive values are removed.

Separate local credentials and server accounts

The Windows client protects its credential file with Windows DPAPI. The server stores account records, including a password hash. Local activity logs and queued events are separate files and are not encrypted by MemoryRaven.

Context sized for a conversation

Replies are capped, so context is less likely to swamp the conversation you are actually trying to have. The summary comes first, recent detail after, and your assistant is told when to ask for a shorter window.

API keys to copy
None
Installer
Not available yet
Assistant access
Read-only
If you go offline
Bounded queue
Computer access
Scoped per authorization

MemoryRaven is not open for new accounts yet

We are validating the Windows installer before offering downloads or new account access.

Account access is not open yet