Configuration reference
Every setting of the service and of its operator commands
(memini serve, migrate,
bootstrap, reindex, credentials,
workspace llm, eval), with its default and
what a change of it needs; then the variables of the client commands and
of the compose stack.
Where settings come from
Highest precedence first:
- The environment:
MEMINI_and the setting name in upper case, sohttp_portisMEMINI_HTTP_PORT. A variable set to the empty string counts as not set; otherMEMINI_*variables are ignored. - A TOML file, if
MEMINI_CONFIG_FILEholds its path (operator commands also take--config FILE). Keys are setting names at the top level; an unknown key is an error, so a typo does not go unnoticed. - The defaults below.
log_level = "DEBUG"
http_port = 6364
git_local_roots = ["/srv/repos", "/home/me/src"]
git_max_blob_size = "2M"Secrets come from the environment only, as
MEMINI_NAME (the value) or MEMINI_NAME_FILE
(the path of a file that holds it, as Docker and compose secrets provide
it; trailing line breaks are dropped). Setting both is an error, and so
is a secret in the TOML file, which holds nothing secret and can be
shared. No log line, error message or output holds a secret's value.
Lists are TOML arrays in the file, or
comma-separated in the environment
(MEMINI_GIT_LOCAL_ROOTS=/srv/repos,/home/me/src).
Sizes are a number of bytes, or a number with a binary
unit: 512M, 5G, 1GiB (1K = 1024
bytes).
Settings are checked together at start. A missing or invalid one stops the service, or the operator command, with exit code 2 and one line per problem; each line names the setting and never its value.
What a change needs
- restart: the service reads its settings once, at
start. With compose,
docker compose up -drecreates the container when its environment changed. Operator commands read the settings on every run. - restart, then
memini reindex: the setting is part of the index fingerprint:embedding_provider,embedding_model,embedding_dimension,embedding_digest,embedding_document_prefix,embedding_pooling,embedding_max_tokens. After a change, search and indexing in a workspace whose index was built with other values are refused, with a message that names the command, untilmemini reindex --workspace NAMErebuilds that index. - A setting with a procedure of its own says so in the table.
Settings
Database
The Postgres database, with pgvector, that the service and the operator commands use. Backup is a dump of it; the mirrors of git sources are fetched again.
| Setting | Values | Default | Change needs | Description |
|---|---|---|---|---|
database_urlMEMINI_DATABASE_URLMEMINI_DATABASE_URL_FILEsecret |
string, not empty | required | restart | A Postgres connection URL. Compose builds it from
MEMINI_DB_PASSWORD. |
HTTP server and request limits
Where the server listens, and the limits of every HTTP and MCP
request. A body over its limit is refused by its
Content-Length before it is read, or as soon as more of it
arrives (413); the ingest route has its own limit,
ingest_max_body_size. The statement limits apply to the
database statements a request runs; search has its own,
search_timeout. Rate limits: per token, a budget of reads
and one of writes; per client address, one for requests whose token is
not known to be valid. Each budget is a minute's worth of requests,
refilled steadily, and is kept in the process: budgets start full after
a restart.
| Setting | Values | Default | Change needs | Description |
|---|---|---|---|---|
http_hostMEMINI_HTTP_HOST |
string, not empty | 127.0.0.1; in the image: 0.0.0.0 |
restart | The address the HTTP server binds to. |
http_portMEMINI_HTTP_PORT |
integer, 1 to 65535 | 6364; in the image: 6364 |
restart | A TCP port from 1 to 65535. The HTTP API and MCP (/mcp)
listen on it. |
http_max_body_sizeMEMINI_HTTP_MAX_BODY_SIZE |
size, 1K to 1G | 1M |
restart | The largest request body outside the ingest and raw response routes, 1K to 1G: bytes, or a number with K, M or G. |
http_statement_timeoutMEMINI_HTTP_STATEMENT_TIMEOUT |
number, 0 to 86400 | 30 |
restart | Seconds one database statement of a request may run, 0 to 86400; 0: the database's default. |
http_idle_in_transaction_timeoutMEMINI_HTTP_IDLE_IN_TRANSACTION_TIMEOUT |
number, 0 to 86400 | 300 |
restart | Seconds a request's transaction may stay idle, 0 to 86400; 0: the database's default. |
http_rate_limitMEMINI_HTTP_RATE_LIMIT |
boolean | true |
restart | True or false: limit requests per token and per client address. |
http_rate_limit_readsMEMINI_HTTP_RATE_LIMIT_READS |
integer, 1 to 1000000 | 600 |
restart | Reads (GET, and every MCP request but a call of a write tool) per minute per token, 1 to 1000000. |
http_rate_limit_writesMEMINI_HTTP_RATE_LIMIT_WRITES |
integer, 1 to 1000000 | 120 |
restart | Writes (POST, PUT, PATCH, DELETE, and MCP calls of write tools) per minute per token, 1 to 1000000. |
http_rate_limit_unauthenticatedMEMINI_HTTP_RATE_LIMIT_UNAUTHENTICATED |
integer, 1 to 1000000 | 60 |
restart | Requests per minute per client address whose token is not known to be valid, 1 to 1000000. |
Logging
The service logs one JSON object per line, on stderr.
| Setting | Values | Default | Change needs | Description |
|---|---|---|---|---|
log_levelMEMINI_LOG_LEVEL |
one of CRITICAL, ERROR,
WARNING, INFO, DEBUG |
INFO |
restart | One of CRITICAL, ERROR, WARNING, INFO, DEBUG. The lowest level the service logs; any case. |
log_slow_statement_thresholdMEMINI_LOG_SLOW_STATEMENT_THRESHOLD |
number, 0 to 86400 | 1 |
restart | Seconds a database statement may take before it is logged as slow, 0
to 86400; 0: no log. A statement of the service over it is logged at
WARNING with its duration, the request or run it served, and its text
with every literal replaced by ?, never its
parameters. |
Metrics
GET /metrics serves counts and timings in the Prometheus
text format, to a scraper that sends metrics_token as its
bearer token; without the token the route does not exist. A Memini token
is not accepted there. No metric is labelled by a workspace, source,
stream or person; memini_base_current says what
GET /v1/status?gate=true says.
| Setting | Values | Default | Change needs | Description |
|---|---|---|---|---|
metrics_tokenMEMINI_METRICS_TOKENMEMINI_METRICS_TOKEN_FILEsecret |
string, 32 to 512 characters | unset | restart | A random secret of at least 32 characters, without whitespace, e.g.
from openssl rand -hex 32. Unset: there is no
/metrics route. Prometheus sends it with
authorization: {credentials_file: ...}; it reads counts
only. |
Operator
v1 has one user. The operator commands act as one principal, named here.
| Setting | Values | Default | Change needs | Description |
|---|---|---|---|---|
principalMEMINI_PRINCIPAL |
string | owner |
nothing to restart (operator commands only); keep it after the first
memini bootstrap |
1 to 63 lowercase letters, digits, '-' or '_', starting with a
letter or digit. The principal the operator commands act as, created on
first use: memini bootstrap gives it the first admin token,
so it owns what is made with that token. |
Embeddings
Off until embedding_provider is set: the service starts
without them, but does not sync git sources, and search refuses with a
message that names the setting. openai_compatible sends the
text of chunks and queries to the API at embedding_url
(Ollama, llama.cpp, vLLM or a cloud service): which endpoint, on this
machine or not, is your explicit choice. local runs an ONNX
model in the process; it needs the optional dependencies
memini[local-embeddings] and pins the model files by
embedding_digest, which is verified before they are loaded.
A setting of the provider not chosen, or any embedding_*
setting without a provider, is an error.
With embedding_provider = openai_compatible and nothing
else set, these defaults apply
(ollama pull qwen3-embedding:0.6b first):
embedding_url:http://127.0.0.1:11434/v1embedding_model:qwen3-embedding:0.6b- and while the model is
qwen3-embedding:0.6b:embedding_dimension:1024embedding_digest:ac6da0dfba84a81fdbfbaf330198c33cd77c4cdfc53e8bc50eb581914a15621dembedding_query_prefix:Instruct: Given a search query, retrieve the relevant code or documentation that can help answer the query\nQuery:
A value that is set wins. Another model gets no dimension, digest or
query prefix from here: it needs its own
embedding_dimension.
| Setting | Values | Default | Change needs | Description |
|---|---|---|---|---|
embedding_providerMEMINI_EMBEDDING_PROVIDER |
one of openai_compatible, local |
unset | restart, then memini reindex |
openai_compatible or local. Unset: embeddings are off. |
embedding_modelMEMINI_EMBEDDING_MODEL |
string, 1 to 200 characters | unset; with openai_compatible:
qwen3-embedding:0.6b; required with local |
restart, then memini reindex |
A model name; for local, a Hugging Face repository such as org/name. |
embedding_dimensionMEMINI_EMBEDDING_DIMENSION |
integer, 1 to 16000 | unset; with openai_compatible and its default model:
1024; required with any provider |
restart, then memini reindex |
The model's vector size, an integer from 1 to 16000. |
embedding_digestMEMINI_EMBEDDING_DIGEST |
string, 1 to 200 characters | unset; with openai_compatible and its default model:
see above; required with local |
restart, then memini reindex |
For local, sha256: and 64 lowercase hex digits; for
openai_compatible, the version of the served model, such as its Ollama
digest. For local it is verified before the model files are
loaded. For openai_compatible it is recorded, not checked:
after ollama pull brings a new build, set its digest. |
embedding_batch_sizeMEMINI_EMBEDDING_BATCH_SIZE |
integer, 1 to 2048 | 32 |
restart | Texts per batch, an integer from 1 to 2048. |
embedding_timeoutMEMINI_EMBEDDING_TIMEOUT |
number, more than 0, at most 3600 | 60 |
restart | Seconds per request, more than 0, at most 3600. |
embedding_query_prefixMEMINI_EMBEDDING_QUERY_PREFIX |
string, at most 200 characters | empty; with openai_compatible and its default model:
see above |
restart | At most 200 characters, e.g. 'query: '. Put before every search query before it is embedded. |
embedding_document_prefixMEMINI_EMBEDDING_DOCUMENT_PREFIX |
string, at most 200 characters | empty | restart, then memini reindex |
At most 200 characters, e.g. 'passage: '. Put before every chunk before it is embedded. |
embedding_urlMEMINI_EMBEDDING_URL |
string, at most 500 characters | unset; with openai_compatible:
http://127.0.0.1:11434/v1 |
restart | openai_compatible only. The base URL of an
OpenAI-compatible API, e.g. http://127.0.0.1:11434/v1 (the
default). It receives the text of every chunk and query. |
embedding_api_keyMEMINI_EMBEDDING_API_KEYMEMINI_EMBEDDING_API_KEY_FILEsecret |
string, not empty | unset | restart | openai_compatible only. Sent to
embedding_url in the Authorization
header. |
embedding_model_dirMEMINI_EMBEDDING_MODEL_DIR |
path | unset; required with local |
restart | local only. A directory for the model files;
created if missing. The files of each digest get a directory of their
own in it. |
embedding_model_fileMEMINI_EMBEDDING_MODEL_FILE |
string, at most 200 characters | onnx/model.onnx |
a new embedding_digest, which covers the file's path;
so restart, then memini reindex |
local only. The path of the ONNX file in the
model repository, e.g. onnx/model.onnx. The digest covers this file and
tokenizer.json, with their paths. |
embedding_revisionMEMINI_EMBEDDING_REVISION |
string | unset | restart | local only. A commit of the model repository,
40 lowercase hex digits. Missing model files are fetched at this commit;
without it, they must be in place. |
embedding_poolingMEMINI_EMBEDDING_POOLING |
one of mean, cls |
mean |
restart, then memini reindex |
local only. mean or cls. How the token vectors
of a text become one: their mean, or the vector of the first token. |
embedding_max_tokensMEMINI_EMBEDDING_MAX_TOKENS |
integer, 1 to 32768 | 512 |
restart, then memini reindex |
local only. An integer from 1 to 32768. A
longer text is cut to this many tokens. |
Search
The limits of one search, and how a hybrid search fuses its
retrievers: each contributes search_candidates results,
merged by reciprocal rank fusion with k search_rrf_k, each
retriever's ranks weighted. The reranker (what it gains and costs:
docs/requirements.md, section 9, Reranker) is none by
default and local in the image, which carries its runtime
and its files, about 610 MB; they are verified against
search_reranker_digest each time the service starts, before
they are loaded. Outside the image, local needs the
optional dependencies memini[local-embeddings], and its
files are fetched at the pinned revision when the service first starts
with it. A search_reranker_* setting while
search_reranker is none is an error, but for
search_reranker_model_dir. Another model or model file
needs its own digest, and another model its own revision; in the image
it also needs a writable search_reranker_model_dir, such as
a volume, as the image's own is read-only.
The pinned default uses per-token INT8 on Linux arm64 and on Linux x86-64 where every CPU reports AVX-512 VNNI. Other Linux x86-64 CPUs use per-channel Q7 weights, including when CPU flags are unavailable. Both graphs read the same verified weights from the cache. Other platforms and model digests use the supplied model graph. Graph provenance
| Setting | Values | Default | Change needs | Description |
|---|---|---|---|---|
search_timeoutMEMINI_SEARCH_TIMEOUT |
number, more than 0, at most 600 | 5 |
restart | Seconds one search statement may run, more than 0, at most 600. |
search_candidatesMEMINI_SEARCH_CANDIDATES |
integer, 1 to 1000 | 50 |
restart | Candidates per retriever of a hybrid search, an integer from 1 to 1000. |
search_rrf_kMEMINI_SEARCH_RRF_K |
integer, 1 to 1000 | 1 |
restart | The k of reciprocal rank fusion, an integer from 1 to 1000. |
search_dense_weightMEMINI_SEARCH_DENSE_WEIGHT |
number, more than 0, at most 100 | 3 |
restart | The weight of the dense retriever's ranks in the fusion, more than 0, at most 100. |
search_lexical_weightMEMINI_SEARCH_LEXICAL_WEIGHT |
number, more than 0, at most 100 | 1 |
restart | The weight of the lexical retriever's ranks in the fusion, more than 0, at most 100. |
search_rerankerMEMINI_SEARCH_RERANKER |
one of local, none |
none; in the image: local |
restart | local: a cross-encoder run in the process; none: no reranking. The
model loads in the background at start; until it is loaded, and if it
cannot load, search keeps the fused order, and the log says why once.
Under compose, MEMINI_SEARCH_RERANKER=none in the settings
file turns it off. |
search_rerank_depthMEMINI_SEARCH_RERANK_DEPTH |
integer, 1 to 100 | 3 |
restart | How many of the best fused hits the reranker reorders, an integer
from 1 to 100. Hits below it keep their fused place.
memini eval --reranker reads it too, so it may be set while
search_reranker is none. |
search_reranker_modelMEMINI_SEARCH_RERANKER_MODEL |
string, 1 to 200 characters | n24q02m/Qwen3-Reranker-0.6B-ONNX |
restart | A Hugging Face repository such as org/name. |
search_reranker_model_fileMEMINI_SEARCH_RERANKER_MODEL_FILE |
string, at most 200 characters | onnx/model_yesno_quantized.onnx |
restart | The path of the ONNX file in the model repository. The digest covers
this file and tokenizer.json, with their paths. |
search_reranker_digestMEMINI_SEARCH_RERANKER_DIGEST |
string | sha256:eaa6edfadd2440e4f80851b3ff7701f26e8ec6aacaa2fedcee8ce86afd09afea |
restart | sha256: and 64 lowercase hex digits, the digest of the model files. Files that do not match are never loaded. |
search_reranker_revisionMEMINI_SEARCH_RERANKER_REVISION |
string | 342a8a54de23bba03eb8cce5f5bef503edf18315 |
restart | A commit of the model repository, 40 lowercase hex digits. Missing model files are fetched at this commit. |
search_reranker_model_dirMEMINI_SEARCH_RERANKER_MODEL_DIR |
path | unset; in the image: /opt/memini/models |
restart | An absolute directory path for the model files, created if missing; unset: $XDG_CACHE_HOME/memini/models, or ~/.cache/memini/models. In the image it holds the default model's files, read-only. |
search_reranker_max_tokensMEMINI_SEARCH_RERANKER_MAX_TOKENS |
integer, 64 to 8192 | 1024 |
restart | The most tokens of one query and hit put to the model, 64 to 8192. A longer prompt is cut at the end of the hit. Each token costs time: about 0.5 ms on an Apple M4 Max. |
search_reranker_timeoutMEMINI_SEARCH_RERANKER_TIMEOUT |
number, more than 0, at most 600 | 5 |
restart | Seconds the reranking of one search may take, more than 0, at most 600. Past it, the search answers in the fused order and logs a warning. |
search_reranker_threadsMEMINI_SEARCH_RERANKER_THREADS |
integer, 0 to 256 | 0 |
restart | Threads of one model run, 0 to 256; 0: as many as the CPUs or a CPU
limit of the container allow, at most 4. onnxruntime would take one per
core, and each step of a run waits for the slowest thread: on a machine
shared with other work, fewer threads answer sooner. The log line
reranking with ... says how many. |
Git connector
The service keeps a bare mirror of every git source and reads code from it. Remote sources use https or ssh; plain http and local paths are refused unless allowed here. Credentials are global for now: one https token and one ssh key.
| Setting | Values | Default | Change needs | Description |
|---|---|---|---|---|
git_mirror_dirMEMINI_GIT_MIRROR_DIR |
path | unset; in the image: /var/lib/memini/mirrors |
restart; mirrors are not moved: the next sync of each source fetches its mirror anew | An absolute directory path; unset: $XDG_DATA_HOME/memini/mirrors, or ~/.local/share/memini/mirrors (the image sets /var/lib/memini/mirrors). |
git_fetch_timeoutMEMINI_GIT_FETCH_TIMEOUT |
integer, 1 to 86400 | 600 |
restart | Seconds a clone or fetch may take, 1 to 86400. |
git_max_mirror_sizeMEMINI_GIT_MAX_MIRROR_SIZE |
size, at least 1 | 5G |
restart | The largest size of one mirror on disk: bytes, or a number with K, M, G or T. |
git_max_blob_sizeMEMINI_GIT_MAX_BLOB_SIZE |
size, at least 1 | 1M |
restart | The largest file read from a mirror: bytes, or a number with K, M, G or T. |
git_allow_httpMEMINI_GIT_ALLOW_HTTP |
boolean | false |
restart | True or false: allow plain http:// remotes (no TLS). |
git_egress_proxyMEMINI_GIT_EGRESS_PROXY |
string | unset | restart | HTTP(S) CONNECT proxy origin for restricted Git egress; the proxy must enforce destination address and port ACLs; unset: legacy transports. |
git_allowed_originsMEMINI_GIT_ALLOWED_ORIGINS |
list of strings | empty | restart | Exact HTTPS origins permitted through git_egress_proxy,
comma-separated; empty with a proxy: deny network Git; requires
git_egress_proxy. |
git_local_rootsMEMINI_GIT_LOCAL_ROOTS |
list of paths | empty | restart | Absolute directory paths, comma-separated, under which local repositories may be sources; empty: no local sources. With compose, mount the directory into the container as well; compose.yaml has an example. |
git_https_tokenMEMINI_GIT_HTTPS_TOKENMEMINI_GIT_HTTPS_TOKEN_FILEsecret |
string | unset | restart | A token for https remotes, without whitespace. |
git_https_usernameMEMINI_GIT_HTTPS_USERNAME |
string | x-access-token |
restart | The user name sent with git_https_token: printable
ASCII, no spaces. |
git_https_token_hostsMEMINI_GIT_HTTPS_TOKEN_HOSTS |
list of strings | empty | restart | Host names, comma-separated, that git_https_token is
sent to; required when the token is set. |
git_ssh_key_pathMEMINI_GIT_SSH_KEY_PATH |
path | unset | restart | The absolute path of the private key for ssh remotes. |
git_ssh_known_hosts_pathMEMINI_GIT_SSH_KNOWN_HOSTS_PATH |
path | unset | restart | The absolute path of the known_hosts file that ssh remotes are checked against. |
Ingest
The limits of one batch of records sent to the ingest API: the request body, the records in it, and the payload of one record as stored.
| Setting | Values | Default | Change needs | Description |
|---|---|---|---|---|
ingest_max_body_sizeMEMINI_INGEST_MAX_BODY_SIZE |
size, 1K to 1G | 16M |
restart | The largest request body of an ingest batch or of an agent run's raw response, 1K to 1G: bytes, or a number with K, M or G. |
ingest_max_recordsMEMINI_INGEST_MAX_RECORDS |
integer, 1 to 100000 | 1000 |
restart | The most records and tombstones in one ingest batch, 1 to 100000. |
ingest_max_payload_sizeMEMINI_INGEST_MAX_PAYLOAD_SIZE |
size, 1K to 1G | 1M |
restart | The largest payload of one record as stored, 1K to 1G: bytes, or a number with K, M or G. |
Credential store
The API keys and OAuth tokens of sources are stored encrypted
(AES-256-GCM) with credential_key, which never enters the
database. To rotate it: make a new key
(openssl rand -base64 32), set it as
credential_key and move the old one to
credential_old_keys, and restart; run
memini credentials rotate-key; when it reports nothing
left, remove credential_old_keys and restart.
| Setting | Values | Default | Change needs | Description |
|---|---|---|---|---|
credential_keyMEMINI_CREDENTIAL_KEYMEMINI_CREDENTIAL_KEY_FILEsecret |
string | unset | restart with the old key in credential_old_keys, then
memini credentials rotate-key |
32 random bytes in base64, e.g. from
openssl rand -base64 32. Unset: the service starts, and
storing or reading a credential fails with a message that names this
setting. |
credential_old_keysMEMINI_CREDENTIAL_OLD_KEYSMEMINI_CREDENTIAL_OLD_KEYS_FILEsecret |
list of strings | empty | restart; remove them once memini credentials rotate-key
reports nothing left |
Keys still read during a rotation, comma-separated, each 32 random
bytes in base64, e.g. from openssl rand -base64 32; only
together with credential_key. |
MCP service executor
The service syncs mcp sources on its own with the built-in MCP
client, and connects only where these settings allow: an endpoint inside
mcp_allowed_endpoints (https, unless
mcp_allow_http), whose host resolves to no internal address
(loopback, private, link-local, the cloud metadata address) unless its
network is in mcp_allowed_networks, checked again at every
connection and redirect; a stdio server only from an approved command or
invocation. Every list is empty by default, so no mcp source reaches
anything until it is listed. A source's OAuth token endpoint is an
endpoint too.
| Setting | Values | Default | Change needs | Description |
|---|---|---|---|---|
mcp_allowed_endpointsMEMINI_MCP_ALLOWED_ENDPOINTS |
list of strings | empty | restart | Hosts (mcp.example.com, host:port, *.example.com) or URL prefixes
(https://mcp.example.com/v1),
comma-separated, that mcp sources may connect to; empty: none. Checked
when an mcp source is added or changed, and before every connection; a
change applies to the next sync. With compose, a server on the host is
host.docker.internal, whose address is internal: list its
network as well. |
mcp_allowed_networksMEMINI_MCP_ALLOWED_NETWORKS |
list of strings | empty | restart | Networks in CIDR notation, comma-separated, of internal addresses (loopback, private, link-local) that MCP endpoints may resolve to all the same; empty: none. |
mcp_allow_httpMEMINI_MCP_ALLOW_HTTP |
boolean | false |
restart | True or false: allow plain http:// MCP endpoints (no TLS). |
mcp_allowed_commandsMEMINI_MCP_ALLOWED_COMMANDS |
list of paths | empty | restart | Absolute paths, comma-separated, of the executables mcp sources may run as stdio servers; empty: none. The path given is the path allowed; a symlink is not resolved. With compose the executable must exist in the container. |
mcp_allowed_invocationsMEMINI_MCP_ALLOWED_INVOCATIONS |
list of StdioInvocations | empty | restart | Approved stdio invocations with command, args and optional
credential_env; a JSON array in the environment or an array of tables in
TOML. For a listed command, the ordered arguments and credential
variable must match one entry exactly, even when mcp_allowed_commands
also lists the command. Example:
[{"command":"/opt/mcp","args":["--ro"],"credential_env":"HEARTH_TOKEN"}]. |
mcp_call_attemptsMEMINI_MCP_CALL_ATTEMPTS |
integer, 1 to 10 | 3 |
restart | Attempts of one call of an mcp sync, the first included, 1 to 10.
Each try is a call of the recipe's [rate] budget. A call
that times out, fails on the network, gets HTTP 429 or a server error,
or answers isError, is tried again after a backoff; one out
of tries leaves its window incomplete. |
LLM providers
At most two OpenAI-compatible chat APIs, declared by the slot they
are set in: llm_local_* is trusted with the content of
every source, llm_cloud_* (https only) only with sources
whose llm_policy is cloud, in workspaces that
opted in (memini workspace llm). The slot is the
declaration; it is never inferred from the URL. Each slot needs its URL
and its model. No provider is configured by default, and then nothing is
extracted.
| Setting | Values | Default | Change needs | Description |
|---|---|---|---|---|
llm_local_urlMEMINI_LLM_LOCAL_URL |
string, at most 500 characters | unset | restart | The base URL of an OpenAI-compatible API you trust with every
source's content, e.g. http://127.0.0.1:11434/v1. Unset,
and llm_cloud_url unset too: no LLM is called. |
llm_local_modelMEMINI_LLM_LOCAL_MODEL |
string, 1 to 200 characters | unset | restart | The model the local API serves. Required when
llm_local_url is set. |
llm_local_api_keyMEMINI_LLM_LOCAL_API_KEYMEMINI_LLM_LOCAL_API_KEY_FILEsecret |
string, not empty | unset | restart | A key for the local API, without whitespace. Only with
llm_local_url. |
llm_local_structured_outputMEMINI_LLM_LOCAL_STRUCTURED_OUTPUT |
one of json_schema, json_object,
none |
json_schema |
restart | json_schema, json_object or none: how the local API is asked for
JSON. Only with llm_local_url. The reply is validated
against the schema in every case. |
llm_cloud_urlMEMINI_LLM_CLOUD_URL |
string, at most 500 characters | unset | restart | The https:// base URL of an OpenAI-compatible API that may read
sources whose llm_policy is cloud, in workspaces that opted in. Not the
same endpoint and model as llm_local_url. |
llm_cloud_modelMEMINI_LLM_CLOUD_MODEL |
string, 1 to 200 characters | unset | restart | The model the cloud API serves. Required when
llm_cloud_url is set. |
llm_cloud_api_keyMEMINI_LLM_CLOUD_API_KEYMEMINI_LLM_CLOUD_API_KEY_FILEsecret |
string, not empty | unset | restart | A key for the cloud API, without whitespace. Only with
llm_cloud_url. |
llm_cloud_structured_outputMEMINI_LLM_CLOUD_STRUCTURED_OUTPUT |
one of json_schema, json_object,
none |
json_schema |
restart | json_schema, json_object or none: how the cloud API is asked for
JSON. Only with llm_cloud_url. The reply is validated
against the schema in every case. |
llm_timeoutMEMINI_LLM_TIMEOUT |
number, more than 0, at most 3600 | 300 |
restart | Seconds per LLM request, more than 0, at most 3600. Only with
llm_local_url or llm_cloud_url. |
Decision extraction
With an LLM provider, each thread or meeting that ingest stores new
or changed gets an extract job: one LLM call extracts its
decisions and open questions, each with the messages or segments that
say it, and one call per item checks it against the most similar active
items of the workspace (a duplicate, or replacing or answering some). A
source or workspace whose llm_policy is none
is skipped. Without a provider no job is enqueued.
| Setting | Values | Default | Change needs | Description |
|---|---|---|---|---|
extraction_candidatesMEMINI_EXTRACTION_CANDIDATES |
integer, 1 to 50 | 8 |
restart | Active items, 1 to 50, that a new item is checked against for duplicates, replacements and answers. Items whose evidence the provider may not read are left out. A larger number costs more input tokens per check. |
extraction_candidate_retrieverMEMINI_EXTRACTION_CANDIDATE_RETRIEVER |
one of lexical, hybrid |
hybrid |
restart | lexical or hybrid: how those items are found, by the words of the
statements or by their words and their vectors. hybrid
needs an embedding provider and is lexical without one; it
embeds each statement once, with the configured model, and keeps the
vector with its item. |
Other settings
Settings that no group above takes yet.
| Setting | Values | Default | Change needs | Description |
|---|---|---|---|---|
retention_intervalMEMINI_RETENTION_INTERVAL |
integer, 60 to 604800 | 3600 |
restart | Seconds between two retention passes over a source that has retention_days, 60 to 604800. |
retention_batch_sizeMEMINI_RETENTION_BATCH_SIZE |
integer, 1 to 100000 | 1000 |
restart | Records one transaction of a retention pass deletes at most, 1 to 100000. |
Client commands
Client commands (memini workspace create,
source, sync, runs,
token, search and the others that are not
operator commands) are HTTP clients of the running service. They read
these variables from the environment, and nothing else: not the
service's settings, nor its config file.
| Variable | Default | Description |
|---|---|---|
MEMINI_URL |
http://127.0.0.1:6364 |
The service's base URL, http:// or
https://, and TLS is whatever the URL says. A URL with a
user name or password is refused: the token is the credential.
--url overrides it. |
MEMINI_TOKEN |
required, or MEMINI_TOKEN_FILE |
The token, a secret: sent as Authorization: Bearer …,
never printed. No option takes the token itself, which shell history and
the process list would keep. An admin creates one with
memini token create; the first comes from
memini bootstrap. |
MEMINI_TOKEN_FILE |
unset | The path of a file that holds the token; trailing line breaks are
dropped. Setting both it and MEMINI_TOKEN is an error.
--token-file overrides both. |
MEMINI_CLIENT_TIMEOUT |
60 seconds | Positive, finite seconds to wait for HTTP activity. Connection attempts use at most 10 seconds. |
Docker compose
compose.yaml (Docker Compose 2.24.0 or later) takes its
configuration from three places, none of them required. After a change,
docker compose up -d recreates the service.
What compose reads:
.env
compose.yaml reads these variables from the environment,
or from .env next to it (.env.example lists
them), and the paths of the secret files below. They configure compose
on the host; none of them is a setting of the service.
| Variable | Default | Description |
|---|---|---|
MEMINI_HTTP_PORT |
6364 |
The port on the host's 127.0.0.1 the service is published on; MCP at
http://127.0.0.1:6364/mcp by default. Not the setting
http_port: in the container compose sets
MEMINI_HTTP_PORT to 6364, where the service listens. |
MEMINI_DB_PORT |
55432 |
The port on the host's 127.0.0.1 the database is published on. |
MEMINI_DB_PASSWORD |
memini |
The password of the database role, a development default; letters
and digits only, as compose puts it into
MEMINI_DATABASE_URL as it is. Set it before the first
start: the database keeps the password it was created with. |
MEMINI_ENV_FILE |
.env.memini |
The settings file, which compose passes into the service's
container; relative to compose.yaml, or absolute. Optional:
a file that does not exist is skipped. |
MEMINI_RERANKER_FILES |
baked |
The reranker's model files in the image compose builds:
baked, fetched at the build (about 610 MB, once: the layer
cache keeps them) and checked against their pinned SHA-256;
none leaves them out, and the reranker then says in the log
that it is off. With none, set
MEMINI_SEARCH_RERANKER=none in the settings file as
well. |
The settings of the
service: .env.memini
Compose passes every line of .env.memini, or of the file
MEMINI_ENV_FILE names, into the service's container as it
is: one MEMINI_NAME=value per line, the variables of the
settings above. So any setting can be set there, a new one included,
without a change of compose.yaml;
.env.memini.example lists them. Its values show in
docker inspect: a secret goes into a file (below)
instead.
Compose sets these itself, over a line of that file:
MEMINI_DATABASE_URL (the database of the stack, with
MEMINI_DB_PASSWORD), MEMINI_HTTP_HOST
(0.0.0.0), MEMINI_HTTP_PORT
(6364), MEMINI_GIT_MIRROR_DIR
(/var/lib/memini/mirrors). Inside the container 127.0.0.1
is the container itself: with embedding_provider, set
MEMINI_EMBEDDING_URL=http://host.docker.internal:11434/v1
for Ollama on the host. The image has the runtime of in-process models,
for the reranker, but no embedding model and no writable place for one,
so the local embedding provider is not set up there.
Secrets and other files
Each of these variables, set in .env, names a file on
the host: relative to compose.yaml, or absolute. Compose
mounts it read-only into the container as a compose secret and sets the
same variable there to the file's path in the container, where the
service reads it; the file's content never shows in
docker inspect. Unset, compose mounts the empty
secrets/unset instead and leaves the setting unset. The
service runs as uid 10001 and must be able to read the file: on Linux it
keeps its owner and mode in the container, so give it to that user
(sudo chown 10001 FILE, mode 400); Docker Desktop and
OrbStack let the container's user read it. ssh refuses a private key
that others can read: keep the key's mode 600 or 400. Compose does not
notice a change of a file's content:
docker compose up -d --force-recreate memini reads it
again.
| Variable | In the container | Setting |
|---|---|---|
MEMINI_EMBEDDING_API_KEY_FILE |
/run/secrets/embedding_api_key |
embedding_api_key |
MEMINI_GIT_HTTPS_TOKEN_FILE |
/run/secrets/git_https_token |
git_https_token |
MEMINI_GIT_SSH_KEY_PATH |
/run/secrets/git_ssh_key |
git_ssh_key_path |
MEMINI_GIT_SSH_KNOWN_HOSTS_PATH |
/run/secrets/git_ssh_known_hosts |
git_ssh_known_hosts_path |
MEMINI_CREDENTIAL_KEY_FILE |
/run/secrets/credential_key |
credential_key |
MEMINI_CREDENTIAL_OLD_KEYS_FILE |
/run/secrets/credential_old_keys |
credential_old_keys |
MEMINI_LLM_LOCAL_API_KEY_FILE |
/run/secrets/llm_local_api_key |
llm_local_api_key |
MEMINI_LLM_CLOUD_API_KEY_FILE |
/run/secrets/llm_cloud_api_key |
llm_cloud_api_key |
MEMINI_METRICS_TOKEN_FILE |
/run/secrets/metrics_token |
metrics_token |
Local repositories
A local repository is a source only under
git_local_roots, and the container sees no directory of the
host until it is mounted: mount the directory that holds the
repositories, read-only, and list its path in the container in
MEMINI_GIT_LOCAL_ROOTS. compose.yaml has an
example under volumes.