ravnori

curl -fsSL https://github.com/daefery/ravnori/releases/latest/download/install.sh | sh

Why ravnori

Three weeks after an agent changed your code, you want to know why. The chat is gone, and the commit message says "fix tests".

Most memory tools can't help, because they write a summary while you work, through a hook that has to run, and keep little of what the summary came from. When the hook never runs, that session is never saved, and nothing tells you.

ravnori works from the transcripts instead. Claude Code, Codex and pi already save every prompt, command and edit to disk. ravnori reads those files, so:

• No session goes missing. A crash, a skipped hook or a model outage only delays a memory. On the author's machine ravnori re-read 1.45 GB of transcripts from scratch with zero duplicates, and rvn doctor tells you exactly how far behind it is.
• Every answer has a source. Each memory points to the transcript lines it came from. rvn ask "why did we..." or rvn ask "what did we do yesterday" answers from the record and lists what it used, so you can check.
• Old memories say they're old. Each one knows whether the code its session wrote is still in your repository. On ravnori's own repo, 75% of the lines agents wrote last month are still there; memories about the rest say so, so they don't send your agent the wrong way.
• Your agents share one memory. A Codex session hears what Claude Code did in the same project an hour ago.
• You see which lines an agent wrote, even in commits from before you installed it.

It runs on the Claude Code or Codex sign-in you already have. One binary, no server, no Node, Python, Docker or vector database.

Install

One line for Linux (x86_64, arm64), WSL and macOS. No Rust needed.

curl -fsSL https://github.com/daefery/ravnori/releases/latest/download/install.sh | sh

The script downloads the binary for your system, checks its SHA-256 and puts it in ~/.local/bin. Then it connects Claude Code, Codex and pi, backing up each settings file first, and starts a small background service that writes memories with your Claude Code or Codex sign-in. Keep working as usual. After your first session:

rvn doctor        # ends with "status: OK"

The viewer is at http://127.0.0.1:37777.

Using Codex? Codex runs no hook until you trust it. Open Codex, type /hooks and trust ravnori's three hooks. Until then your Codex sessions are still recorded, they just don't get memories back. rvn doctor reminds you.

Used mnem before? ravnori is its new name, and the command is now rvn. Install it the same way: it moves your memory from ~/.mnem to ~/.ravnori and replaces mnem's hooks, tools and service. mnem keeps working as a link to rvn until 0.7.0. What changes.

Try it first, without changing anything

Want to see what ravnori finds before it touches your setup? This installs only the binary and reads your existing transcripts into its own folder. Your agents' settings stay exactly as they are. rvn ask answers with the Claude Code or Codex you are signed in to.

curl -fsSL https://github.com/daefery/ravnori/releases/latest/download/install.sh | RAVNORI_NO_SETUP=1 sh
~/.local/bin/rvn backfill                        # read every saved session
~/.local/bin/rvn ask "what did we do last week"  # answered from them, with sources
~/.local/bin/rvn search "flaky test"             # search all of them, across agents
~/.local/bin/rvn doctor                          # what was found, per agent

To undo it, delete ~/.local/bin/rvn and ~/.ravnori. To keep it, run rvn install --watch.

/plugin marketplace add daefery/ravnori
/plugin install ravnori@ravnori
/ravnori:setup

/ravnori:setup installs the program with the same script. If you also ran the one-line installer, the plugin's copies stay quiet, so nothing runs twice.

git clone https://github.com/daefery/ravnori && cd ravnori
cargo install --path . --locked --features fastembed
rvn install --watch

Needs Rust 1.89 or newer. --features fastembed adds the local model used for meaning-based search.

• RAVNORI_VERSION=v0.6.2 picks a release, RAVNORI_BIN_DIR another folder, and RAVNORI_NO_SETUP=1 installs only the binary.
• Linux with glibc 2.35 to 2.38 (Ubuntu 22.04, Debian 12) and Intel Macs get a "lite" build: the same ravnori, with a smaller model for meaning-based search.
• Windows: use WSL.
• The background service downloads the search model once (about 90 MB).

What you get

Answers about past work, with sources

rvn ask answers only from what it found and lists every source it cited: memories, your pinned facts, what you typed and what your agents replied. If they don't cover the question, it says so.

• Why: "why did we drop the merge button?" It says whose reason it was: when you only said "go" to an agent's suggestion, the answer reads "the agent recommended it because …; you chose it", not a reason you never gave. A later "still open" from you outweighs an older note that calls it decided.
• When: "what did we do yesterday?", "what shipped on 4 October?", "last week", "kemarin". Read in your local time, across every project unless you name one, with shipped work first. It works on the first day too, before any memory is written.
• Your agents ask it too. The MCP tool ask (and pi's ravnori_ask) lets an agent answer "what did we decide about X?" from the record instead of digging through git log.

Memory that shows up on its own

You don't call anything. When a session starts, when you type a prompt and when the agent opens a file, ravnori adds the few memories that match:

ravnori recall: past memories matching this prompt (full text: get_observations([ids]))
#2 summary · 3d ago · Stripe sometimes charges a customer twice when the webhook is retried. Find out why and fix it.
#1 bugfix · 3d ago · Replayed Stripe webhooks no longer create duplicate orders

Opening a file brings the memories about that file, each noting whether its edits are still there. Agents can also search on their own through ravnori's MCP tools, and rvn remember "..." pins a fact every agent sees at session start.

A live view of every agent

Every agent's prompts and memories as they happen. Search them, filter by project, take a backup, or fix an agent that isn't connected with one click. (The screenshots use a made-up project: docs/demo.)

Which lines did an agent write?

rvn trace writes Agent Trace records for your commits: which added lines came from which agent session and model. It covers commits from before you installed anything, because the transcripts already hold every edit. Useful in code review, and for teams with rules about AI-written code.

Build your own tools on it

A local, read-only API serves the whole record, so your scripts skip parsing three transcript formats:

rvn api        # prints the address and token
curl -s -H "Authorization: Bearer $(cat ~/.ravnori/api-token)" \
  "http://127.0.0.1:37777/v1/search?q=webhook+retries"

Sessions, events, memories with their evidence, and search, paged so a tool can sync. See docs/api.md.

How it works

1. Capture. Each agent appends prompts, answers, commands and edits to a transcript file. ravnori reads each file as it grows and saves its place in the same database write, so nothing is skipped or read twice.
2. Distil. After each turn, and in the background for anything missed, your own model turns the session into short memories that cite the lines behind them.
3. Recall. Hooks put the best few memories in front of the agent at the right moment, ranked by words and by meaning with a local model, and never repeated in a session.

Privacy

Everything stays on your machine: the record, the memories, the search model, the viewer and the API. There is no ravnori server.

Secrets are removed before anything is stored: API keys, tokens, passwords in URLs.

Two things leave your machine. To write memories, redacted excerpts go to the model you chose, by default your own Claude Code or Codex sign-in. And the background service downloads its search model once from Hugging Face.

rvn forget deletes a memory, a session or a whole project for good. rvn uninstall disconnects your agents and keeps your data until you delete ~/.ravnori.

Everyday commands

Command	What it does
rvn doctor	Is every session captured? Which agents are connected? Anything to fix?
rvn ask "..."	Answer a question about past work (why, when, what shipped), with sources
rvn search <words>	Search everything captured
rvn remember "..."	Pin a fact every agent sees at session start
rvn forget <id>	Delete a memory, session or project for good
rvn backup / rvn restore	Verified snapshots (taken daily by default); --merge adds a teammate's memory to yours
rvn trace	Agent Trace records for this repository's commits
rvn api	The record API's address and token
rvn install / rvn uninstall	Connect or disconnect your agents

Every command, setting and measurement is in docs/reference.md.

How it compares

	ravnori	Tools that write memories during the session
A hook or model call fails	the memory waits; the session is read again from its transcript	a hook that never runs leaves nothing to recover
Where a memory came from	points to the transcript lines	at best a batch of raw tool calls (claude-mem); most keep nothing
Is its code still there?	checked for each memory	not tracked by the tools we checked
Agents	Claude Code, Codex and pi, one shared memory	some cover more (claude-mem 10+)
What you run	one binary	often Node, Python or a vector database

Thirteen tools compared in detail, with sources: the comparison and roadmap.

FAQ

Rarely enough to notice. On the author's machine over the past two weeks, the prompt hook took 26 ms at the median and 167 ms for the slowest 5%. Memories are written after the turn, in the background.

Yes, if you write memories with your Claude Code or Codex sign-in, which is the default. One request covers about 16,000 characters of a finished session, so a short session takes one and a long one several. All background requests together stop at 100 a day. Sessions past that wait for the next day; none are lost, and rvn doctor tells you. You can use any OpenAI-compatible endpoint instead.

Not as often as they should, yet. Judges rate about two thirds of the memories ravnori shows as helpful, but agents open fewer than 1 in 10. Better ways to present them are being tested and measured now (roadmap). The record, the sources and rvn ask work either way.

rvn import brings in your claude-mem history with its ids. ravnori can also recover sessions claude-mem never saved, from the transcripts still on disk. Try it first with the no-changes steps above.

Run rvn uninstall, then delete ~/.local/bin/rvn. To remove your data too, delete ~/.ravnori.

Contributing

Issues and pull requests are welcome. Changes to recall are measured before they ship (rvn eval --gate); see CONTRIBUTING.md.

License

ravnori is free software under the GNU AGPL v3.0. Use it, change it and share it, at work too. If you distribute a modified ravnori, or run one as a service for others, share your changes under the same license. A commercial license is available for uses the AGPL doesn't suit.