Local-first desktop AI memory that reads foreground text, stores long-term context on-device, and uses it for recall and writing across apps.
Goldfish is a local-first desktop assistant designed to remember what a user has been working on across applications and then use that context when writing or answering questions. It runs on macOS and Windows and reads text from the foreground window through the operating system accessibility layer; the company says it does not take screenshots or record the screen.12 The captured text becomes a local memory that can be searched from Goldfish chat, summarized into automatically generated wiki-style pages for people/projects/topics, and used when drafting or rewriting inside ordinary text fields.3
The interaction model is deliberately lightweight. On macOS the default activation is the right Option key; on Windows it is the right Shift key. Goldfish can draft a reply from the surrounding context, rewrite an existing draft with typed or spoken instructions, answer questions about recent work and provide proactive reminders or briefs.3 That makes it closer to a memory layer underneath normal desktop work than a separate chat destination.
The second important capability is MCP. Goldfish runs a local MCP server and can register itself with Claude Desktop, Cursor and Claude Code. Connected clients can search the same locally stored memory and read activity summaries, so a user can ask Claude about prior work without manually exporting a transcript or knowledge base.4 The MCP server is local/stdio and the documentation says local connections are limited to the user's machine and protected by a per-install token.2
Early user evidence supports the basic value proposition. Product Hunt currently shows 5.0/5 from seven reviews and more than 1,000 launch upvotes; reviewers repeatedly praise cross-app context and the reduction in copy-paste prompting, while asking for clearer visibility into what was remembered and more granular memory controls.5 These are encouraging but very small, self-selected samples—not mature reliability data.
Goldfish is currently free. The June 2026 Terms of Service state that the service is provided free of charge and that paid plans or features may be introduced later with pricing presented before charging.6 The current FAQ likewise describes the product as free during beta.7 This makes the present value proposition unusually attractive: the user can build a local memory and test whether context-aware drafting is actually useful without a subscription commitment.
The unresolved issue is future economics. The settings documentation already contains plan states for Free, Goldfish Pro and Founding members, but no stable public Pro price is currently published.3 A user should therefore treat “free” as the current product state, not a promise about long-term pricing. That matters more than usual for a memory product because its usefulness compounds with time: switching later can have a higher behavioral cost than abandoning a generic chatbot.
Installation is direct from Goldfish for macOS and Windows.8 macOS requires Accessibility permission because Goldfish needs to read screen text and write into text fields; Windows does not require the equivalent special permission. Microphone access is optional and only needed for voice dictation.9 There is no conventional mobile app surface in the current product model, so App Store/Google Play ratings should not be used as proxies for this desktop application.
Privacy architecture is the central reason Goldfish is credible enough to consider despite the sensitivity of its function. The long-term captured memory is stored locally on the computer rather than in a Goldfish cloud database. The documentation says there is no backend/cloud storage for that captured memory and that Goldfish's makers cannot browse it.2 Users can pause capture indefinitely or temporarily, blacklist specific applications or browser domains, and wipe the entire local database. Card numbers, CVVs, Social Security numbers and IBANs are masked at capture time by default.2
“Local-first” does not mean every byte remains on the machine. When the user asks the AI to act, the context needed for that request is transmitted for inference. The August 12 privacy policy says requests pass through Cloudflare to a private Microsoft Azure OpenAI deployment with zero data retention, and says Microsoft does not store the request afterward or train on it.10 This is an important distinction: the persistent memory is local, but task-relevant snippets can leave the device for model processing.
The product also places a meaningful trust burden on capture configuration. Goldfish continuously reads foreground-window text when capture is active; this is exactly how it gains the context that makes it useful. Users working with banking, health, legal, HR or confidential client systems should configure the blacklist before ordinary use rather than assuming that local storage alone eliminates sensitivity. The availability of blacklist, pause and delete controls is a strength, but the correct privacy posture still requires deliberate setup.2
Independent evidence is currently the main limitation. Goldfish launched in June 2026, and there is not yet a broad corpus of professional reviews, enterprise deployments, security assessments or reproducible comparative tests. Product Hunt's seven perfect ratings are too few and too close to launch to support a mature satisfaction score.5 Current first-person reports are useful for identifying the workflow benefit, but they cannot establish long-term memory accuracy, false-context rates, resource consumption or enterprise support quality.
Goldfish is most compelling for people whose work is fragmented across many desktop applications. Sales, recruiting, founders, consultants, lawyers, product teams and other knowledge workers often lose time reconstructing context: finding an earlier discussion, pasting a thread into an assistant, explaining who a person is and then asking for a reply. Goldfish tries to remove that setup cost by remembering the work locally and making the context available directly where the user types.15
It can also be valuable for heavy Claude/Cursor users because local MCP turns the memory into a reusable context source for other AI clients rather than locking it inside one chat UI.4 That makes Goldfish conceptually closer to a personal context infrastructure layer than a conventional writing assistant.
It is a weaker fit for users uncomfortable with continuous foreground-text capture even when storage is local, organizations that cannot send any task context to an external model processor, and teams that need mature admin consoles, audit logs, centralized retention policies or contractual enterprise guarantees today. The product's youth should be treated as a real procurement factor: the privacy design is promising, but product maturity and governance maturity are different questions.
Rating: Coming soon.
Products matched by shared tasks, audience, workflow, product type and capabilities.