Memory your AI agent keeps.

Lynx saves what your agent learns as plain files on your machine. Between sessions, a dream reads it back with hindsight, fixes what went stale, and waits for your yes.

$ npm i -g https://lynx.example/lynx.tgz && lnx init

$LYNX CA at launch

Why Lynx

Every session starts from zero.

  1. 01

    Your agent forgets. You explain the repo, it does good work, the session ends, and everything it learned ends with it. Tomorrow you explain it again.

  2. 02

    Notes alone go stale. Saving facts helps until they pile up. A rule written in March quietly contradicts one from May, and the agent trusts whichever it finds first.

  3. 03

    Lynx sleeps on it. It writes down what matters while the agent works. Between sessions it reads everything back with hindsight, merges what repeats, flags what disagrees, and hands you a diff. Nothing changes until you say yes.

Named after the lynx, the animal that sees clearly in the dark. Most of what your agent knows lives there, between sessions, until something brings it back.

A session three weeks later

Asked once, answered from then on.

Your agent asks its memory before it asks you. Search is plain keyword ranking (BM25) over notes and wiki pages, so every result shows why it matched.

The output on the right is real: the CLI run against a small sample project.

~/shop-api
$ lnx recall "how do I run the tests?" --limit 3

 1    1.38  /testing.md
      Run `make test` from the repo root. Needs
      `DATABASE_URL=postgres://localhost/app_test`
      (set in .env.test) since the Postgres move.
 2    1.30  /ci.md
      GitHub Actions on Node 24. The test job runs
      make test against a Postgres service container.
 3    0.69  n_mwmabwfn370fca
      Seed the test database first: pnpm db:seed --env test

How it works

From one session to a memory that gets better.

Four layers, each a folder you can open: transcripts, notes, a wiki, and the short digest your agent wakes up with.

hooks · stop
01

Your agent works.

Claude Code hooks copy each session into a transcript. Other agents save facts through the MCP server with remember. Known secret shapes (API keys, tokens, passwords, private keys) are replaced with [REDACTED] before anything reaches disk.

recall · BM25
02

Recall finds it.

Every note and page is searchable the moment it is saved. Page titles weigh more, and pages linked from a match come along. No embeddings, no vector database, nothing to host.

dream · proposal
03

A dream reviews it, with hindsight.

A dream reads the new transcripts and notes against the wiki. It corrects facts that changed, merges duplicates and links related pages. The result is a proposal you read as a diff.

/testing.md
  run: make test
- no env var needed+ needs DATABASE_URL (since the Postgres move)
session-start
04

The next session wakes up informed.

When a session starts, your agent gets a short digest: the wiki index and the newest notes. It opens the page it needs instead of asking you again.

Install

One command, then restart your agent.

lnx init finds the agents on your machine and wires each one: Claude Code gets hooks and the MCP server, the others get the MCP server.

  • Shows its plan first with --dry-run.
  • Backs up every config file it edits, and leaves alone any file it cannot parse.
  • One memory space per project folder, so projects never mix.
  • Requires Node 22 or newer.

Hooks at session start and stop, plus the MCP server. Add --prompt-recall to look memory up before every prompt; it hands over notes only when at least two words match.

npm i -g https://lynx.example/lynx.tgz
lnx init --claude-code --prompt-recall

Features

Memory that gets reviewed, not only stored.

A wiki your agent can find its way around.

Short markdown pages, one topic each, linked with [[page]]. An index lists every page in one line, so a new session knows where to look.

~/.lynx/shop-api/wiki/
├─ _index.md
├─ testing.md   → [[ci]]
├─ ci.md
├─ deploy.md    → [[ci]]
└─ style.md

Every change is a version.

Each write keeps the page as it was, with the reason. Read the history and bring back any version with one command.

  • v3dream: needs DATABASE_URL now18 Nov
  • v2dream: merged flaky-tests.md11 Nov
  • v1first dream04 Nov

Dream with a key, or with the agent you run.

With an Anthropic key, lnx dream prints a token estimate and stops at your budget. Without one, your agent does the thinking through dream_prepare and dream_submit, on the plan you already pay for.

Secrets are scrubbed on the way in.

API keys, tokens, passwords, private keys and credentials inside URLs become [REDACTED] before any write. It matches known shapes, so keep secrets out of chats anyway.

Diff, then apply or reject.

Nothing in the wiki changes until you run lnx apply. A page you edited after the dream read it is skipped, never overwritten.

0accounts, servers or databases to run
4memory layers, each a folder
19MCP tools, plus a context resource
30CLI commands

Map

See what your agent knows.

lnx map draws a whole memory as one picture: pages, the notes that belong to them, faded notes dimmed, and a dashed line where two facts disagree. One HTML file on your machine. Below is a sample project; hover or search it.

Checks

It notices what you would miss.

Memory goes stale quietly. Lynx looks for it: facts that disagree, notes nobody uses, pages with no clear source. Conflicts and Strength run the real engine code here in your browser; Trace and Outbound show its real output.

Write two facts. Lynx compares them the way lnx conflicts does: same topic with one side negated, or the same setting with a different value. The next dream is asked to settle every pair it finds.

Free

Free, and your data stays yours.

The CLI and the MCP server are MIT licensed. Memory lives in ~/.lynx as plain text: read it, grep it, back it up, put it in git.

The only thing that ever leaves your machine is a dream you start yourself, sent to the model you chose. Dreamed transcripts are cleared after 30 days with lnx retention, or delete the folder whenever you like.

The software$0
Saving, searching, the digest$0, runs on your machine
A dream with your keyWhat your provider charges; the estimate prints first
A dream through your agentPart of the plan you already have
Accounts, servers, telemetryNone

Verify

Built to be checked, not taken on trust.

A memory tool sees your work, so it should be easy to inspect. Each of these takes about a minute, on your own machine.

Read the code this page runs.

The Map, Conflicts and Strength demos above import the engine's own files, served unchanged. Open them and compare with the package.

Check the package before you install it.

The checksum below is computed by the server from the file it serves right now. Download it, hash it, and the two should match. The tests ship inside.

SHA-256
computing…
Size
…
$ curl -sO https://lynx.example/lynx.tgz
$ shasum -a 256 lynx.tgz
$ tar xzf lynx.tgz && cd package && npm install && npm test

Nothing to connect.

No wallet connection, no account, no sign-up. Nothing on this site asks you to sign a transaction or approve a token. $LYNX is shown by its contract address only, so you can look it up yourself.

A record of everything that leaves.

Saving, search, the digest, conflicts and the map all run offline. A dream you run with your own key is the one call that goes out, and each one is written to a local ledger you can read with lnx outbound. MIT licensed; the license ships in the package.

Questions

Asked before you install.

What is a dream?

A review pass. It reads the transcripts and notes saved since the last dream, compares them with the wiki, and writes a proposal: pages to add, rewrite or remove, each with a reason. You read it with lnx diff and accept it with lnx apply.

Can a dream make my memory worse?

It can propose a bad change, which is why nothing applies on its own. You see every line before it lands, a page you touched in the meantime is skipped, and every write keeps the previous version for lnx restore.

How much does a dream cost?

It prints a token estimate before it sends anything and refuses to go over --budget. For a big backlog, narrow it with --since or --max-inputs. Through your agent it costs nothing extra.

How does search work?

BM25 keyword ranking over notes and pages, with titles weighted higher and linked pages included. A result needs words in common with the query, so the agent is told to try the likely terms.

Where does my data live?

In ~/.lynx/<project>/: notes.jsonl, transcripts/, wiki/ with its .history/, and proposals/. Set LYNX_HOME to put it elsewhere.

Which agents work with it?

Claude Code (hooks and MCP), Codex, Cursor and the Claude desktop app through lnx init, and any client that runs a local stdio MCP server.

Can memory steer my agent somewhere it should not go?

Everything Lynx returns is wrapped and labelled as stored data, and the dream prompt treats transcripts as material to review, never as instructions. Keep reviewing the diffs: that is where you stay in charge.

What does the simulation do?

Recorded mode plays three sessions and a dream in your browser. Live mode lets you talk to a small agent that has Lynx's remember and recall; your messages go through this site's server to the model, and the memory stays in your tab.

Let your agent sleep on it.

Free and MIT licensed. No account, no server, no telemetry. Watch three sessions and one dream, then install it on your own machine.