Learn

This topic gets a memory store running and then keeps it running.

The tutorials are one path, in order. Each ends with a store that does more than the one before it, and every command shown is a command you can run.

1. Install memhtml and initialize a store: npm i -g memhtml, then memhtml init. One package carries the whole system and installs two binaries; the page also covers building from a clone, which is what contributors do.
2. Write your first memory: memhtml write, the file it commits into the git tree, and why one fact goes in one file.
3. Retrieve it: memhtml search and memhtml recall, and what separates them.
4. Wire up the MCP server: memhtml serve mcp, and the tools and resources a client sees.

The operations pages are task-shaped how-tos, one per section of the runbook. They assume a store already exists, and each answers a question you arrived with.

Page	Answers
Configure the environment	Which variables the binary reads, and what each one degrades when it is absent
Initialize a store	Why a fresh clone runs memhtml init before its first merge
Wire up your coding agent	What integrations install writes per host, and how uninstall removes exactly that
Hooks and recall	What each hook injects, why a failing hook prints nothing, and how to run one by hand
Run the store day to day	The daily verbs, the cron lines, and what moves the access plane
Share one store between a CLI and a server	Whether a command and a running server can touch one database at once
Rebuild the index	When update is not enough, and how to clear a vector-space mismatch
Preserve the state plane	The one set of facts the git tree cannot reproduce
Run and review a sleep cycle	Seventeen phases on a branch, how to read them, when the merge refuses
Check the discrimination gate	The gated number that says whether retrieval still tells a fact from its own negation
Audit and publish the corpus	Every memhtml doctor finding and its fix
Index session transcripts	The trace plane, and the firewall between it and memory retrieval
Diagnose poor retrieval	Where to look when nothing errors and the answers are wrong
Recover from a lost index	What a clone plus a rebuild restores, and what it cannot

Before you start

Every command writes exactly one JSON envelope to stdout and nothing else, and sends its logs to stderr. Exit 0 is success, 2 is a usage error you fix by changing the call, and 1 is a runtime failure you fix by changing the repository or the environment (apps/cli/src/envelope.ts:87). So every command on this site is safe to pipe into jq and safe to run from cron, and the examples show the envelope rather than describing it.

The git tree is the system of record, and .memhtml/index.db is a projection of it: delete the index and you lose time rather than memories. That property is why the operations pages read the way they do. Most recovery is a rebuild, and most of what looks like corruption is a stale watermark. Store layout and path algebra develops it.

If you are an AI agent rather than a person reading a page, start with memhtml manifest. It answers with every command, flag, response type, error code, and environment variable this binary accepts, and it answers on a machine with no repository, no database, and no credentials. Come back here for the worked paths.