One of them holds your working life on someone else's servers. The other is a folder on your disk. Here is the honest version of that trade, including the parts where the other side wins.
Strip away the branding and every one of these tools has the same three jobs. It is worth naming them, because the differences between products are entirely in how they do each one.
A plain CLAUDE.md that you maintain by hand does the third job
well and the first two badly, which is why it is stale within a fortnight.
Everything below is about the other two.
A hosted context layer solves capture by connecting to the places your work already lives. You grant OAuth access to Slack, Gmail, Notion, your calendar, your repos. It ingests continuously, builds a profile on its servers, and serves it back to your AI tools over a remote MCP endpoint.
This is a genuinely good design and it wins on coverage. It sees the meeting you did not write down. It works on your phone, your laptop and your work machine identically. If your context genuinely lives in email threads and Slack DMs, nothing local is going to match it, and you should use one.
The price of that design is structural, not a flaw to be patched:
Unforgetter takes the opposite side of that trade. The memory is a folder of markdown on your Mac. The MCP server is a small binary talking to your agent over stdio on your own machine. There is no account because there is no server to have an account on.
Capture happens where you already work rather than through connectors: the agent writes a line to today's log when something is decided, and a session-end pass catches what it missed. Distillation runs once a day, spawning your logged-in CLI to fold the day's logs into a single current-state file, then archiving what it consumed.
That last detail is the one people miss. Because consolidation runs through the CLI you are already paying for, the intelligence in the loop costs us nothing to provide, which is why there is no subscription. A hosted layer cannot copy this. It has to run the model itself, so it has to bill you monthly.
| Hosted context layer | Unforgetter | |
|---|---|---|
| Your memory lives | On their servers | In a folder on your Mac |
| To read it | Their web app, or an export | Any text editor. It is markdown |
| Coverage | Slack, Gmail, Notion, calendar, repos | What you and your agent do |
| Credentials held | One OAuth token per connected source | None |
| Cost | Monthly, per seat, forever | Free. Consolidation uses your own CLI login |
| Across machines | Automatic | Your problem - but it is a folder, so git or Syncthing |
| If the company dies | Export and migrate | Nothing happens. It is your folder |
| Version history | Whatever they expose | Whatever git gives you |
A hosted memory has to own every connection it offers you. That is why those products are, in practice, a race to add integrations: coverage is the product, so every source you care about has to be built, maintained and paid for by them.
A local memory does not have that problem, because it is not the only thing your agent can reach. You have probably already wired up an MCP server or two - GitHub, a database, your notes, whatever you needed. Your agent reads from those directly. Unforgetter's job is only to remember what came of it.
We do not integrate with Slack, and we are not going to. Your agent can already get there on its own, using a server you chose and trust. What it could not do before was remember any of it tomorrow.
The comparison above is not free of costs on our side, and pretending otherwise would be exactly the kind of landing-page noise we are trying to avoid.
If most of your context arrives through email, meetings and chat, and you would rather pay monthly than think about it, use a hosted layer. That is a reasonable trade and we are not going to pretend it is not.
If your work happens in a terminal, if the thing you keep re-explaining is your own projects and decisions, and if you would rather your working life not sit in someone else's database - this one was built for you.