memhouse is the memory of your coding agents. It makes “tell me what my agents did last summer” work.
Your agents have amnesia: every session starts from zero, and the old ones are thrown away. memhouse is their memory palace: every session of every agent, kept whole, not a byte lost, in a database you own. Any agent can walk in and recall what it did, or what another agent did. No more copy-pasting to carry the context: it is all already there, intact, waiting to be used.
Your agents are deleting your history right now. Claude Code deletes a session after 30 days by default (cleanupPeriodDays), and a rewind, an edit or a /clear takes messages out of other agents’ stores. A longer setting keeps the files on one laptop. memhouse turns them into memory every agent can use.
It keeps
The sessions of 12 coding agents, captured as they run: Claude Code, Codex, Cursor, Cursor CLI, VS Code, Zed, Gemini CLI, Copilot CLI, OpenCode, Goose, Devin and agy. Every prompt, reply, tool call and its full output.
It gives it back
A plugin hands it to your agents: /mem:recall answers from your history, /mem:access shares it and reads what others shared, /mem:ask counts it, /mem:house says what is kept, /mem:admin runs the house.
It stays yours
Your own ClickHouse, on your machine or your server. No memhouse cloud, no account, no telemetry. You choose who else reads it: a session, a project, a folder.
npm install -g memhouse && memhouse onboard
onboard keeps what is still on your disk today, then every session as it runs.
no memhouse cloudno accountno telemetryyour database
Why memhouse exists
We believe the work you do with agents is knowledge.
That knowledge is priceless to your agents, so it should be kept safe.
It belongs to you, not to whichever agent was open.
You explained the constraints. Your agent tried three approaches. Together you found the fix. Then the session ended, and the tool that held it started counting down to deleting it. That work should be kept whole, owned by you, readable by every agent you use, and shared only with the people you choose.
01 / The loss
cleanupPeriodDays). Then the transcript is deleted from your disk.It is not the only way history goes. Edit a message in Cursor and the ones after it are deleted. /clear in Goose, a revert in OpenCode, a rewind in Antigravity: each takes rows out of the agent's own store. What survives sits in a dozen private formats, on whichever laptop you happened to use.
The fix you found in March is only worth something if you can still find it. memhouse reads your agents' session files as they grow, every five minutes by default, and keeps what it reads, including messages the agent later removes. It cannot bring back what was deleted before you installed it. Start while your history is still there.
On the machine memhouse was built on, the oldest Claude Code transcript still on disk was nine days old.
02 / Recall
memhouse keeps the record itself, in full: not a summary, not an embedding. Nothing is truncated, tool output included, so the agent reading it sees what actually happened, not what someone guessed was worth keeping.
Ask in words, search the text, or count in SQL.
/mem:recall searches past sessions, reads the transcripts and answers with citations, from inside Claude Code; Codex has its own memhouse skill.memhouse search finds an exact error string across every session, every machine, every agent.memhouse resume prints the command that reopens a past session in the editor it came from, in the right folder.You measured it in a Codex session on 2 October: a 4% hit rate, and two stale-price bugs traced to invalidation. You replaced it with a materialized view and deleted cache.ts.
codex · 2026-10-02 · ~/shop/checkout · my-laptop · memhouse resume codex:5e0c…
memhouse finds the sessions; your agent reads them and answers. memhouse runs no model of its own.
Everything the agent did, kept.
And underneath, every line your agents wrote, archived verbatim, so a session can be rebuilt from the archive alone, including what later went away.
It runs on your machine (memhouse dashboard), reads your house, and sends nothing anywhere.
03 / Every agent
Use the agent that fits the task, not the one that happens to hold the context. Ask Codex about a session you had in Claude Code last week; let Cursor read the approaches you already tried. Every agent's sessions land in the same house, so a new tool or a new machine does not mean starting over, and you never have to choose one agent.
Claude Code gets /mem:recall, /mem:ask, /mem:house, /mem:access and /mem:admin; Codex gets a skill and a note in its AGENTS.md. Any other agent can use the CLI, or a system-prompt snippet from memhouse prompt.
Sessions reach the house while they run, every five minutes by default (you can set it shorter), sub-agents included. One agent can read what another has been doing, or what its sub-agents found.
Install it on the desktop and the laptop, and both ship into the same house. Each row carries the machine it came from, so a question can span both or stay on one.
Captured as you work · twelve agents · plus claude.ai exports
04 / Together
Keep a team's sessions on one ClickHouse server. Each person's tables are private until they share them. Share a project's history, and a colleague's agent can search the sessions behind the code: the failed attempts, the commands that worked, the decisions as they were made. Nobody stops to write the wiki. The work already wrote it.
The decisions behind the code live in the sessions that made them. Shared, they are one question away, on day one.
The on-call agent searches every session shared with it for the exact error string, and finds the night someone fixed it, if anyone has.
A reviewer's agent reads the author's session, not just the diff: what was tried, what failed, and why this approach won. Only if the author shared it.
One agent's research is the next one's starting point, across people, machines and tools, as far as the shares reach. The second agent starts where the first one stopped.
Each member's tables are theirs alone until they share them. The server stamps who wrote every row, and a fingerprint says which machine it came from. No client can fake either.
06 / Yours
memhouse runs no service of its own and no AI in its pipeline. A small program on your machine parses your agents' files and writes rows to the ClickHouse you choose. Everything after that is a database doing what databases do.
A container on this laptop (memhouse deploy --local, bound to loopback), a server you run, ClickHouse Cloud, or your team's server. Rows go nowhere else.
There is no memhouse service to sign up for. Parsing happens on your machine; status and doctor ask the npm registry whether a newer version is out.
Embeddings turn stack traces, flags and numbers into approximate neighbours. memhouse stores the exact rows, with no model in its write or read path; the agent reading them does the thinking.
The server stamps the database user on every row as it is written, and each row carries a fingerprint of the machine it ran on, not a hostname a client could set.
A message an agent deletes stays as a removed version. Raw lines are archived. A session can be rebuilt from the archive alone.
Apache-2.0. Read every line that touches your sessions, run it, change it.
Install today, and the sessions still on your disk become the oldest ones you will never lose.
npm install -g memhousememhouse onboardfinds your agents, sets up a house, ships what is on disk/mem:recall what was I working on last week?Node 24+ · macOS and Linux · or hand it to your agent: memhouse prompt --install
07 / Questions
memhouse keeps what your agents saw, so a secret an agent printed is in your house, exactly as it was already in the agent's own files. That is why the house is a database you own, where you decide who has an account and can read every row yourself, why each member's tables are theirs alone until shared, why a share can be one project or one session, and why a share never reaches the raw archive. Share on purpose: a revoke cannot un-read what was read. And there is no command yet to forget one session; for now, removing data from the house is a database operation.
Your sessions go only to the ClickHouse you point it at, which can be a container on the same laptop (memhouse deploy --local, bound to loopback). There is no memhouse cloud, no account and no telemetry. memhouse status and doctor ask the npm registry whether a newer version is out.
Each pass, every five minutes by default, reads only what changed since the last one. Nothing is truncated, so the house grows with your work; it is ClickHouse, so plain SQL shows exactly what each table holds and how much room it takes.
Captured as they work: Claude Code, Codex, Cursor (the editor and cursor-agent), VS Code (with Insiders, Code-OSS and VSCodium) including its Copilot chat, GitHub Copilot CLI, Gemini CLI, the Antigravity CLI, Devin's local CLI, Goose, OpenCode and Zed. A claude.ai data export can be imported. Not yet on a new house: Kiro, Codebuff, Command Code, Copilot for JetBrains, the Antigravity IDE and Devin's desktop app.
Most memory tools decide what is worth keeping and store it as summaries or embeddings. memhouse keeps everything, as typed rows, and adds no model of its own. Search matches the exact text, so the same query finds the same error string every time, and the same rows answer "what did we spend on this model last month". It also reads the history already on your disk on the day you install it.
memhouse is free and open source. The database runs wherever you put it: a container on your laptop, a server you already run, or ClickHouse Cloud.
Yes. One ClickHouse server holds a house per member. An admin runs memhouse invite alice; Alice installs with the file she is given and sets her own password. Nothing is shared until someone runs memhouse share. Start with the docs.
Node 24 or newer, on macOS or Linux, and either a ClickHouse to point at or Docker or Podman for memhouse deploy --local. npm install -g memhouse compiles nothing and needs no flags.