Independent AI intelligence Two editions daily · ET
Fervor AI

Analysis · November 20, 2026 · repo

LingerSimplexAI Labslocal-aiagent-memoryprivacy

Linger Puts a Personal AI Agent on Your PC. Whether Your Inbox Stays There Depends on One Setting.

The open-source agent reads your mail every night and runs a 125B model on a 12 GB GPU. Here is how to set it up so "local" actually means local.

The most useful personal AI agent is also the most invasive one. To remind you about the client who is waiting on a quote, it has to read your mail. To prepare you for Thursday's meeting, it has to read your calendar, your Slack and your notes. Every product in this space eventually asks for all of it.

Most of them then send it to someone else's servers.

Linger, an open-source agent from SimplexAI Labs that sat at #2 on Trendshift's daily board on November 20, makes the opposite promise. Its README opens with "a Claude-class brain running locally on your PC," and its local inference doc says "every layer lives on your machine." Linger README, Local inference

Reading the docs closely turns up the condition behind that promise. Linger keeps your inbox on your machine when its agents run on the local model. Its nightly memory and wiki jobs use whatever runtime each agent is placed on (the runtime is the program that runs the agent and calls the model), and if that runtime talks to a remote endpoint, the endpoint "receives the input needed for the task."

So the one decision that matters most is the model source. Choose Local model before you connect Gmail.

What Linger is

Linger is a desktop app (Mac, Windows and Linux) plus a small server, built around one main agent the project calls Another Me. It is young: the repository's first commit is dated November 10, 2026, and its desktop releases went from v0.1.3 to v0.1.6 between November 18 and November 20. This piece reviews tag desktop-v0.1.6.

Three ideas carry the project.

A big model on a normal GPU. Linger's inference engine, PolyStrata, runs mixture-of-experts models. A mixture-of-experts model is split into many specialist sub-networks and only uses a few of them for each word, so Linger keeps the busiest specialists on your graphics card and parks the rest in system memory or on SSD. That is how the README gets the 125B Qwen3.8-Flash-Next model onto "a 12 GB card with 64 GB of RAM." Linger README

A wiki of your working life. Connect your inbox and Another Me "reads back through the last few months," then each night files mail, chats and meetings into Markdown pages: one page per client, contact, project and topic, and one per meeting. Every agent you run can search those pages. Wiki and memory

An agent that speaks first. Another Me checks Gmail, Outlook, Google and Outlook Calendar and Slack mentions every 10 minutes, decides whether something deserves your attention, and messages you, schedules a follow-up, or stays silent. Before an interview, pitch or client meeting it gathers your notes, recordings, mail and web results into a prep document. Proactive Another Me

The part the headline does not say

Linger supports three model sources, and each computer serves all of its agents from one source at a time:

  • Local model: a model on your own machine, run through Pi, an open-source coding agent, with Linger's local gateway in between.
  • Subscription: Claude Code or Codex, using the plan you already pay for.
  • BYOK (bring your own key): your own OpenAI-compatible or Anthropic endpoint and key, through the local gateway.

Local inference

Now the important paragraph. Under "Background Memory and Wiki," the doc says memory facts, agent memory and "personal Wiki generation use the Runtime on the Computer where the relevant Agent is placed." On Pi with a local model, those calls stay on the machine, and the doc says they "never switch to Codex or Claude Code." On a Codex or Claude Code placement, they use that runtime's default endpoint, and "if that endpoint is remote, it receives the input needed for the task." The wiki doc repeats the point for agent memory: extraction runs "with its Codex or Claude Code and the user's subscription." Local inference, Wiki and memory

In plain words: the files stay on your disk either way. But the work of reading your mail and turning it into wiki pages is done by a model, and if that model lives in someone's cloud, your mail goes there to be read.

That is a reasonable design. It is not what "every layer lives on your machine" suggests on a first read.

There is a second, smaller leak path. On the Local model and BYOK sources, agents search the web through Linger's own commands, which use whichever search providers the server is set up with (Tavily, Brave Search, Firecrawl, Exa or You.com) and fall back to DuckDuckGo with no key. A local agent preparing your client meeting can still send search queries about that client to a search company.

The project seems to be steering toward the local path itself. The setting that switches model sources later offers Local model or BYOK, not Subscription, and a commit tagged desktop-v0.1.4 on November 19 is titled "hide the Subscription model source tab." Commit history

Put this into practice

1. Check your hardware before you install anything. The README's hardware table maps each Mac and NVIDIA tier to a recommended model. The README's prose shows the 125B model on a 12 GB NVIDIA card with 64 GB of RAM, while its tier table lists that model on a 12 GB card with 32 GB and notes it is not the default pick for that machine. A 16 GB Mac gets a much smaller 4B model. Linger's installer detects your GPU, memory and bandwidth and picks the strongest model that fits by its own scoring, then only counts it installed after it answers a real question.

2. Install the desktop app or the server. The desktop app is the easy path on macOS 14 or later (Apple silicon only), Windows x64 or Linux x64. The macOS build is unsigned, so the README has you clear the quarantine flag first:

xattr -dr com.apple.quarantine /Applications/Linger.app

For the server route, Docker Compose works (give Docker's VM at least 5 GiB of memory), though the Compose server runs no models by itself:

git clone https://github.com/simplexai-labs/linger.git
cd linger
git checkout desktop-v0.1.6   # the version reviewed here
cp .env.example .env
docker compose up --build --detach --wait

The README's owner-account example uses the password test1234, and it tells you to change it before anyone else can reach the server. Do that before step 4.

3. Choose Local model during setup, and load it. This is the step that decides where your mail goes. Pick Local model, let the installer verify it, and confirm it is loaded: the doc notes that under Local model, "downloaded is not loaded." If you also run Claude Code or Codex on the same computer for coding, remember that background jobs follow the agent's placement, so keep the personal-assistant agents on the local runtime.

4. Connect one source, not five. Start with a single inbox. Let it run one night.

5. Read what it wrote. The wiki pages are plain Markdown, and the app keeps a Change Log on each page with one dated line per rewrite, edit, retraction or merge. Facts cite the message they came from. Open a few client pages and check them against the real threads. If the summaries are wrong, you want to know before you trust its reminders.

6. Decide on web search deliberately. If search queries about your clients matter to you, pick a provider whose terms you accept, or treat the prep feature as something to use for public topics only.

7. Expect the first week to be noisy. Another Me learns from how you reply. Each night a process the project calls Dreaming reads how its messages landed and turns repeated feedback, like "tell me sooner" or "don't remind me about this kind," into rules. Repeated replies of the same kind become a rule, and naming the kind outright makes one straight away. The rules stay on your machine with the rest of your memory.

Honest limitations

  • The speed numbers are the company's own lab figures. The README's footnote says decode speeds were measured "mostly on a simulated tier," a larger card held to the tier's memory budget, "so a real machine of the tier can be slower." The 95% score on LoCoMo, a benchmark of questions about long conversations, is also the project's own claim.
  • The license is not plain Apache 2.0. The Linger Open Source License adds conditions: you need a commercial license to offer Linger to people outside your organization as a hosted service, you may not remove the logo or copyright from its frontend, and contributors agree the producer can make the license stricter or looser later. Fine for personal and internal use; read it before building a product on top. License
  • It is ten days old. Four desktop releases in three days is great for momentum and hard on stability. Back up anything you would hate to lose and expect breaking changes.
  • Search is keyword-only. The wiki search uses PostgreSQL full-text search with no embedding model, plus up to six alternate wordings the agent supplies. It is simple and predictable, but it will miss matches that share meaning without sharing words.
  • No Intel Mac desktop app. The macOS package is Apple silicon only; the README points Intel Macs to Docker Compose.
  • Local means slow first tokens on big models. The README's own table lists a first token "after about 3 minutes" for the 125B model on a 12 GB Linux tier after a 62K-token prompt.

Why this matters beyond Linger

Every personal agent will face the same split. The files, the memory and the interface can all live on your laptop, while the reading and summarizing happens wherever the model runs. "Local" in a README describes where things are stored. Where your data is read depends on the model.

Linger is unusually honest about this in its docs, which is why it was easy to find. Most products in this space never write the sentence down.

If you want an agent that knows your clients without a vendor knowing them too, Linger is worth an evening. Pick the local model first, connect one inbox, and read the pages it writes before you let it speak for you.

Sources: Linger README, Local inference, Wiki and memory, Proactive Another Me, License, commit history, Trendshift. Reviewed at tag desktop-v0.1.6 (commit 77abdcb, November 20, 2026).