Configuration

Environment variables for the server and the clients.

The server is configured with environment variables (or the matching m3m0s serve flags). With Docker Compose, put them in .env.

Server

Variable	Default	Purpose
`OPENAI_API_KEY`	none (keyword-only)	Key for embeddings (`text-embedding-3-small`). Without it the server runs keyword-only: full-text search and exact-match dedup.
`M3M0S_OPENAI_BASE_URL`	OpenAI	Alternative OpenAI-compatible embeddings endpoint.
`M3M0S_PORT`	`19234`	HTTP port.
`M3M0S_DATA_DIR`	`~/.m3m0s/data` (`/data` in the image)	Where the SQLite database lives.
`M3M0S_HOST_PORT`	`19234`	Host port published by `compose.yaml`.
`M3M0S_ADD_SYNC_TIMEOUT`	`300ms`	How long `add` waits before answering `202` and finishing in the background.
`GOGC`	`50`	Go GC target. Lower values trade CPU for a lower memory peak.
`GOMEMLIMIT`	not set	Optional soft memory limit, for example `600MiB`.

Fact extraction and chat (optional)

Extraction turns one input into several short atomic facts. It stays off until you set a key.

Variable	Default	Purpose
`M3M0S_EXTRACT_API_KEY`	none (off)	Enables LLM fact extraction.
`M3M0S_EXTRACT_BASE_URL`	`https://openrouter.ai/api/v1`	OpenAI-compatible endpoint for extraction.
`M3M0S_EXTRACT_MODEL`	`deepseek/deepseek-v4-flash`	Extraction model.
`M3M0S_CHAT_API_KEY`	extraction key	Enables the WebUI chat.
`M3M0S_CHAT_BASE_URL`, `M3M0S_CHAT_MODEL`	OpenRouter, `deepseek/deepseek-v4-flash`	Chat provider and model.

Background jobs

Variable	Default	Purpose
`M3M0S_RETRO_EXTRACT_INTERVAL`	`2h`	Retroactive extraction job interval.
`M3M0S_FORGET_SWEEP_INTERVAL`	`12h`	Forgetting sweep interval.
`M3M0S_RETENTION_MODEL`	`legacy`	Score the sweep uses: `legacy` (one 30-day half-life) or `per-kind` (half-life by nature of fact, opt-in).
`M3M0S_REEMBED_BATCH`, `M3M0S_REEMBED_RPS`	`64`, `5`	Batch size and request rate of the re-embedding job.
`M3M0S_JOBS_TICK`	`30s`	Scheduler heartbeat.
`M3M0S_INSIGHTS_CPU_PERCENT`	`60`	CPU duty cycle of the memory-health analysis.
`M3M0S_INSIGHTS_MAX_MINUTES`	`30`	Wall-clock limit of that analysis.

Ingestion filter

Automatic capture drops or tags tooling boilerplate before storing it.

Variable	Default	Purpose
`M3M0S_FILTER_MODE`	`default`	Global mode: `default`, `off`, `tag` or `drop`.
`M3M0S_FILTER_MODES`	none	Per-rule modes that win over the global one, for example `diff-trivia-fact=drop`.
`M3M0S_FILTER_RULES`	none	JSON file of rules that replace, add or disable built-in ones.

Clients

These are read by the CLI, the MCP bridge, the watcher and the hooks.

Variable	Default	Purpose
`M3M0S_URL`	`http://localhost:19234`	Server URL.
`M3M0S_API_KEY`	none	API key.
`M3M0S_HOOK_DOMAIN`	derived from the working directory	Domain captured memories are filed under.

Recall tuning variables are on the Recall and tuning page.

Embedding provider at runtime

The embedding model, endpoint, dimensions and key can also be changed without a restart, from the WebUI (Embedding page) or the CLI:

m3m0s embedding show
m3m0s embedding estimate --model text-embedding-3-large --dimensions 1024
echo "$NEW_KEY" | m3m0s embedding set --model text-embedding-3-large --api-key-stdin
m3m0s embedding reembed start --watch

The key is never accepted as an argument and never printed. If a change leaves stored vectors in another model or size, set shows how many and an approximate cost, and asks you to confirm.

Edit this page on GitHub