Configuration variables
This page lists every environment variable that RememberStack's engine and
the remember package read. Configuration
explains how they reach the containers.
The variables in Client are read by the remember CLI, Python client and
MCP server on your own machine. Every other section applies to a self-hosted engine.
The shipped compose.yaml hands every variable in .env to the engine
containers. It sets three itself, overriding .env:
REMEMBERSTACK_DATABASE_URL, REMEMBERSTACK_MINIO_ENDPOINT_URL and the
container's REMEMBERSTACK_SELFHOST_API_PORT
(Configuration).
Booleans accept true/false (also 1/0). An empty value counts as
unset for optional settings.
Client
| Variable | Default | Type | Purpose |
|---|---|---|---|
REMEMBER_API_URL | http://127.0.0.1:8000 | URL | The deployment endpoint. |
REMEMBER_API_KEY | unset | secret | The API key, raw or as Bearer …. |
REMEMBER_PROJECT | unset | id or name | The project to use when the key covers several. |
REMEMBER_CONFIG_DIR | $XDG_CONFIG_HOME/remember or ~/.config/remember | path | Where credentials.json lives. |
REMEMBERSTACK_INTERNAL_OPS | false; true in the engine image | bool | Enables remember ops, which works only inside the engine image. Also read as REMEMBER_INTERNAL_OPS. |
REMEMBERSTACK_MCP_INGEST_ROOTS | empty | JSON array or comma-separated paths | Directories from which the local MCP server may ingest a file by path. Empty refuses path ingest. |
REMEMBERSTACK_MCP_PATH_READ_MAX_BYTES | 268435456 | int > 0 | Largest file the local MCP server reads for path ingest. |
XDG_CONFIG_HOME | unset | path | Standard base for REMEMBER_CONFIG_DIR and for agent configuration that remember setup writes. |
APPDATA | unset | path | Windows base for agent configuration that remember setup writes. |
The CLI and the Python client read REMEMBER_API_KEY, REMEMBER_API_URL and
REMEMBER_PROJECT the same way: a flag or argument wins over the variable,
and the variable wins over the credential file
(CLI).
Compose file only
Read by compose.yaml from .env, not by the engine.
| Variable | Default | Type | Purpose |
|---|---|---|---|
REMEMBERSTACK_POSTGRES_USER | required | string | PostgreSQL user; also used to build REMEMBERSTACK_DATABASE_URL. |
REMEMBERSTACK_POSTGRES_PASSWORD | required | string | PostgreSQL password. .env.example ships none; generate it at install. |
REMEMBERSTACK_POSTGRES_DB | required | string | PostgreSQL database name. |
REMEMBERSTACK_POSTGRES_MAX_CONNECTIONS | 300 | int | Read by Compose only. PostgreSQL max_connections. The default stack's ceiling is 231 connections; see Database connections. |
REMEMBERSTACK_BUILD_REVISION | empty | string | Build argument when Compose builds the app image; stamped into it and reported by GET /deployment as build_revision. |
Deployment and API
| Variable | Default | Type | Purpose |
|---|---|---|---|
REMEMBERSTACK_DATABASE_URL | required | SQLAlchemy URL | PostgreSQL connection, postgresql+psycopg://user:password@host:5432/db. Compose builds it from the REMEMBERSTACK_POSTGRES_* values. |
REMEMBERSTACK_SELFHOST_DEPLOYMENT_ID | required | UUID | The deployment's identity; .env.example ships none, generate it at install. Fixed after the first start. |
REMEMBERSTACK_SELFHOST_DEPLOYMENT_SLUG | local | string | Short name. Fixed after the first start. Default Sentry environment. |
REMEMBERSTACK_SELFHOST_DEPLOYMENT_NAME | Local memory | string | Display name. Fixed after the first start. |
REMEMBERSTACK_SELFHOST_DEFAULT_LANGUAGE | en | string | Default language of the deployment. Fixed after the first start. |
REMEMBERSTACK_SELFHOST_API_HOST | 0.0.0.0 | string | Address the API listens on inside the container. |
REMEMBERSTACK_SELFHOST_API_PORT | 8000 | int 1–65535 | Port inside the container (Compose sets 8000). In .env it is the host port Compose publishes. |
REMEMBERSTACK_SELFHOST_API_PUBLISH_ADDRESS | 127.0.0.1 | host address | Read by Compose only. Host address Compose publishes the API port on. Set a token before changing it; see Opening the API to other machines. |
REMEMBERSTACK_SELFHOST_INGEST_BODY_MAX_BYTES | unset (no cap) | int > 0 | Largest POST /ingest body; larger uploads get 413. |
REMEMBERSTACK_SELFHOST_BROWSER_ORIGINS | empty | comma-separated https origins; http only on localhost, 127.0.0.1 or [::1] | Browser origins allowed by CORS. |
REMEMBERSTACK_SELFHOST_TRUSTED_PRINCIPAL_SOURCE | false | bool | Believe X-Ingest-Principal-* headers on POST /ingest. |
REMEMBERSTACK_SELFHOST_API_ADMISSION_KEY_PER_MINUTE | unset (no limit) | int ≥ 0; 0 = no limit | Optional. Requests per minute per credential (bursts of a quarter of it). Over it: 429 rate_limited. |
REMEMBERSTACK_SELFHOST_API_ADMISSION_KEY_IN_FLIGHT | unset (no limit) | int ≥ 0; 0 = no limit | Optional. Requests running at once per credential. Over it: 429 concurrency_limited. |
REMEMBERSTACK_SELFHOST_API_ADMISSION_DEPLOYMENT_PER_MINUTE | unset (no limit) | int ≥ 0; 0 = no limit | Optional. Requests per minute for the whole deployment (bursts of a quarter of it). |
REMEMBERSTACK_SELFHOST_API_ADMISSION_DEPLOYMENT_IN_FLIGHT | unset (no limit) | int ≥ 0; 0 = no limit | Optional. Requests running at once for the whole deployment. |
Storage
| Variable | Default | Type | Purpose |
|---|---|---|---|
REMEMBERSTACK_MINIO_ENDPOINT_URL | required | URL | S3-compatible endpoint. Compose sets http://object-store:8333. |
REMEMBERSTACK_MINIO_ACCESS_KEY | required | secret | Object-store access key; also the bundled SeaweedFS store's admin access key. .env.example ships none; generate it at install. |
REMEMBERSTACK_MINIO_SECRET_KEY | required | secret | Object-store secret key; also the bundled SeaweedFS store's admin secret key. .env.example ships none; generate it at install. |
REMEMBERSTACK_MINIO_REGION | us-east-1 | string | S3 region name used for signing. |
REMEMBERSTACK_MINIO_CONNECT_TIMEOUT_S | 5 | finite seconds, > 0 | Connection timeout for each S3 request. |
REMEMBERSTACK_MINIO_READ_TIMEOUT_S | 30 | finite seconds, > 0 | Socket read timeout for each S3 request. Multipart uploads send at most 8 MiB per part. |
REMEMBERSTACK_SELFHOST_RAW_BUCKET_NAME | remember-raw | string | Bucket for original files. Fixed after the first start. |
REMEMBERSTACK_SELFHOST_ARTIFACTS_BUCKET_NAME | remember-artifacts | string | Bucket for converted Markdown and artifacts. Fixed after the first start. |
REMEMBERSTACK_SELFHOST_CORPUSFS_BUCKET_NAME | remember-corpusfs | string | Bucket for live corpus folders and navigation. Fixed after the first start. |
REMEMBERSTACK_SELFHOST_PROJECTION_WORK_ROOT | /var/lib/rememberstack/projection-work | path | Working directory created by setup. |
REMEMBERSTACK_SELFHOST_FORGET_MANIFEST_ROOT | /var/lib/rememberstack/forget-manifests | path | Where hard-forget manifests are read. |
REMEMBERSTACK_SELFHOST_FORGET_MANIFEST_BUCKET_NAME | unset | string | Keep hard-forget manifests in this dedicated bucket instead of the manifest root. Never one of the three data buckets. |
REMEMBERSTACK_SELFHOST_MIGRATION_CONFIG | alembic.ini | path | Alembic configuration used by setup. |
Authentication
| Variable | Default | Type | Purpose |
|---|---|---|---|
REMEMBERSTACK_SELFHOST_API_BEARER_TOKEN | unset | secret | Shared secret with write scope. Setting it turns authentication on. |
REMEMBERSTACK_SELFHOST_API_BEARER_BIND | unset | <deployment-uuid>:<sha256-hex> | The shared secret's digest instead of the secret. Must match the token if both are set. |
REMEMBERSTACK_SELFHOST_REQUIRE_API_AUTH | false | bool | Refuse to start unless the token, the bind or the key issuer is set. |
REMEMBERSTACK_SELFHOST_API_KEY_ISSUER | unset | URL | The key issuer, matched exactly against iss. Setting it turns signed keys on and requires the tenant id and both URLs below. |
REMEMBERSTACK_SELFHOST_API_KEY_TENANT_ID | unset | string | The issuer's tenant id for this deployment. Keys carry aud = org:<this> and org = <this>. |
REMEMBERSTACK_SELFHOST_API_KEY_PROJECT_ID | the deployment id | string | The issuer's project id for this deployment, matched against a key's projects. |
REMEMBERSTACK_SELFHOST_API_SIGNING_KEYS_URL | unset | http(s) URL | Where the issuer's JWKS of Ed25519 public keys is fetched. |
REMEMBERSTACK_SELFHOST_API_REVOCATION_URL | unset | http(s) URL | Where the issuer's signed revocation document for this deployment is fetched. |
REMEMBERSTACK_SELFHOST_API_KEY_REFRESH_S | 60 | float > 0 | Seconds between fetches of the key set and the revocation document. |
REMEMBERSTACK_SELFHOST_API_REVOCATION_MAX_AGE_S | 3600 | float > 0 | Maximum age of the accepted revocation document; past it every signed key is refused. |
See Authentication and scopes.
Processing layers
Which parts of the pipeline the deployment runs and serves. Both are on by
default; see Processing layers.
setup applies a change.
| Variable | Default | Type | Purpose |
|---|---|---|---|
REMEMBERSTACK_SELFHOST_CLAIMS_LAYER | true | bool | Extract, embed and serve claims. Turning it off also requires REMEMBERSTACK_SELFHOST_FACTS_LAYER=false. |
REMEMBERSTACK_SELFHOST_FACTS_LAYER | true | bool | Build and serve entities, facts, the entity graph and knowledge pages. Requires the claims layer. |
REMEMBERSTACK_SELFHOST_LAYER_SEED_BATCH_SIZE | 500 | int ≥ 1 | Documents setup queues per batch when it starts processing a layer that was switched on. |
Pools and workers
| Variable | Default | Type | Purpose |
|---|---|---|---|
REMEMBERSTACK_SELFHOST_WORKER_RATE_PER_S | 20.0 | float > 0 | Claims per second per worker process. |
REMEMBERSTACK_SELFHOST_WORKER_BURST | 20.0 | float ≥ 1 | Claim burst per worker process. |
REMEMBERSTACK_SELFHOST_WORKER_FALLBACK_POLL_S | 5.0 | float > 0 | Poll interval when no notification arrives. |
REMEMBERSTACK_SELFHOST_WORKER_SESSION_S | 3600.0 | float > 0 | Length of one worker session. |
REMEMBERSTACK_SELFHOST_RETRIEVAL_DATABASE_BUDGET_S | 120.0 | finite float > 0, max 3600 | One retrieval operation’s database deadline, including admission and nested reads. Combined context has two independently bounded children; allow their sum plus overhead in client/proxy timeouts (e.g. MemoryClient(..., timeout=300)). Expiry is an incomplete boundary, not evidence that nothing exists. Larger values hold a read slot and snapshot longer. |
REMEMBERSTACK_SELFHOST_RETRIEVAL_POOL_SIZE | 4 | int 1–32 | Connections for retrieval in the API. |
REMEMBERSTACK_SELFHOST_RETRIEVAL_POOL_TIMEOUT_S | 1.0 | float, max 30 | Wait for a retrieval connection. |
REMEMBERSTACK_SELFHOST_RETRIEVAL_MAX_CONCURRENCY | 4 | int 1–32, ≤ pool size | Retrievals at once. |
REMEMBERSTACK_SELFHOST_GRAPH_POOL_SIZE | 4 | int 1–32 | Connections for graph queries in the API. |
REMEMBERSTACK_SELFHOST_GRAPH_POOL_TIMEOUT_S | 1.0 | float, max 30 | Wait for a graph slot. |
REMEMBERSTACK_SELFHOST_GRAPH_MAX_CONCURRENCY | 2 | int 1–32, ≤ pool size | Graph queries at once. |
REMEMBERSTACK_SELFHOST_GRAPH_WORK_MEM_KIB | 16384 | int 64–65536 | PostgreSQL work_mem per graph query. |
REMEMBERSTACK_SELFHOST_QUERY_ROLE_HOST | unset (the database URL's host) | hostname | Where query_sql connects as the deployment's query role, for example PostgreSQL itself when the database URL points at a connection pooler. |
REMEMBERSTACK_SELFHOST_QUERY_ROLE_PORT | unset (the database URL's port) | int 1–65535 | Port for that connection. |
REMEMBERSTACK_SELFHOST_QUERY_ROLE_SSLMODE | unset (libpq default) | disable, allow, prefer, require, verify-ca, verify-full | TLS mode for that connection. |
REMEMBERSTACK_SELFHOST_QUERY_ROLE_SSLROOTCERT | unset (libpq default) | path | CA certificate file that verifies the server for that connection. |
REMEMBERSTACK_WORK_RETRY_BACKOFF_BASE_S | 2.0 | float | First retry delay; doubles per attempt. |
REMEMBERSTACK_WORK_RETRY_BACKOFF_MAX_S | 60.0 | float | Longest retry delay. |
REMEMBERSTACK_WORK_LEASE_S | 60.0 | float | How long a worker's hold on an item lasts without renewal. A running worker renews it every third of this; an item whose worker died is retried once it runs out (Operating). |
REMEMBERSTACK_WORK_BUDGETS | [] | JSON list | Spend ceilings per deployment, stage and lane (Operating). |
REMEMBERSTACK_E1_EMBED_BATCH_SIZE | 64 | int 1–512 | Chunks per embedding request. |
REMEMBERSTACK_P1_EMBED_BATCH_SIZE | 64 | int 1–1024 | Claims or facts per embedding request. |
Model provider
| Variable | Default | Type | Purpose |
|---|---|---|---|
REMEMBERSTACK_OPENROUTER_API_KEY | required | string | OpenRouter key for every model call. |
REMEMBERSTACK_TYPESAFE_BACKUP | none | none, openai_decisions | Optional availability backup for the Jev adjudication engine. |
REMEMBERSTACK_OPENAI_DECISIONS_API_KEY | unset | secret | Native OpenAI key, required only when the backup is enabled; OPENAI_API_KEY is also accepted outside Compose. |
REMEMBERSTACK_OPENAI_DECISIONS_MODEL | gpt-6-luna | gpt-6-luna | Currently supported native Decisions model. |
REMEMBERSTACK_OPENAI_DECISIONS_BASE_URL | https://api.openai.com/v1 | HTTPS URL | Native endpoint; URLs containing credentials are rejected. |
REMEMBERSTACK_OPENAI_DECISIONS_TIMEOUT_S | 30.0 | positive float | Timeout for the single backup request. |
REMEMBERSTACK_OPENAI_DECISIONS_INPUT_USD_PER_MILLION | 0.10 | positive decimal | Global short-context ordinary-input cost basis. |
REMEMBERSTACK_OPENAI_DECISIONS_CACHED_INPUT_USD_PER_MILLION | 0.01 | positive decimal | Cache-read cost basis. |
REMEMBERSTACK_OPENAI_DECISIONS_CACHE_WRITE_USD_PER_MILLION | 0.125 | positive decimal | 30-minute cache-write cost basis. |
REMEMBERSTACK_OPENAI_DECISIONS_LONG_CONTEXT_THRESHOLD | 272000 | positive integer | Long-context rates apply above this reported input-token count. |
REMEMBERSTACK_OPENAI_DECISIONS_LONG_CONTEXT_MULTIPLIER | 2 | decimal ≥1 | Multiplier on all long-context input categories. |
REMEMBERSTACK_OPENAI_DECISIONS_PRICING_MULTIPLIER | auto | positive decimal or auto | Global 1 or native regional 1.1 automatically; custom endpoints require an explicit factor. |
Native Decisions response handling has only been tested with offline mock
transports. Run a small live API check before enabling this default-off backup.
| REMEMBERSTACK_OPENROUTER_BASE_URL | https://openrouter.ai/api/v1 | URL | OpenRouter API address. |
| REMEMBERSTACK_OPENROUTER_TIMEOUT_S | 120.0 | float > 0 | Per-request timeout. |
| REMEMBERSTACK_OPENROUTER_MAX_COMPLETION_TOKENS | 32000 | int ≥ 1 | Output allowance per chat call, reasoning included. |
| REMEMBERSTACK_OPENROUTER_CHAT_PROVIDER_ORDER | unset | comma-separated or JSON list | Preferred chat providers, with fallback. |
| REMEMBERSTACK_OPENROUTER_CHAT_PROVIDER_ONLY | unset | comma-separated or JSON list | Only these chat providers. Alone it disables fallback; combine with an order naming the same unique slugs to allow fallback within the set. |
| REMEMBERSTACK_OPENROUTER_EMBEDDING_PROVIDER_ORDER | unset | comma-separated or JSON list | Preferred embedding providers, with fallback. |
| REMEMBERSTACK_OPENROUTER_EMBEDDING_PROVIDER | unset | string | The one embedding provider, no fallback. The order wins if both are set. |
| REMEMBERSTACK_OPENROUTER_ZDR | false | bool | Only zero-data-retention endpoints for chat. |
| REMEMBERSTACK_OPENROUTER_REASONING_EFFORT | unset | none, minimal, low, medium, high, xhigh, max | Reasoning effort for every chat call. |
| REMEMBERSTACK_OPENROUTER_REASONING_EFFORT_MAP | unset | JSON object | Reasoning effort per model id; wins over the global value. |
| REMEMBERSTACK_OPENROUTER_CHAT_THROTTLE_RETRIES | 3 | int ≥ 0 | Retries of a chat call after HTTP 429, outside the work attempts. |
| REMEMBERSTACK_OPENROUTER_CHAT_UPSTREAM_OVERLOAD_MAX_RETRY_AFTER_S | 30.0 | float > 0 | Longest wait before retrying an overloaded provider. |
| REMEMBERSTACK_OPENROUTER_INVALID_COMPLETION_CAPTURE_DIR | unset | absolute path | Debug only: keep schema-invalid model answers as files. |
| REMEMBERSTACK_TYPESAFE_API_KEY | unset | string | TypeSafe AI key, required with the jev fact engine. Also read as TYPESAFE_AI_API_KEY. |
| REMEMBERSTACK_TYPESAFE_MODEL | jev-latest | string | TypeSafe model. |
| REMEMBERSTACK_TYPESAFE_BASE_URL | https://api.typesafe.ai/v1 | URL | TypeSafe API address. |
| REMEMBERSTACK_TYPESAFE_TIMEOUT_S | 30.0 | float | TypeSafe request timeout. |
Model seats
See Models and providers.
| Variable | Default | Type | Purpose |
|---|---|---|---|
REMEMBERSTACK_STRUCTURER_MODEL | openai/gpt-5.6-luna | model id | Structure fallback: section anchors when headings are unusable. |
REMEMBERSTACK_STRUCTURER_MAX_PROMPT_CHARS | 200000 | int ≥ 1000 | Largest prompt sent to the structure fallback. |
REMEMBERSTACK_STRUCTURER_MIN_HEADING_DENSITY_PER_10K | 0.25 | float ≥ 0 | Structure check: minimum headings per 10,000 characters. |
REMEMBERSTACK_STRUCTURER_MAX_OVERSIZED_LEAF_RATIO | 0.75 | float 0–1 | Structure check: largest share of oversized sections. |
REMEMBERSTACK_STRUCTURER_MIN_BLOCKS_FOR_LLM | 8 | int ≥ 1 | Accepted for compatibility; not used. |
REMEMBERSTACK_SKELETON_CHECK_MODEL | z-ai/glm-4.7-flash | model id | Checks a proposed section outline. |
REMEMBERSTACK_ROLE_MODEL | z-ai/glm-4.7-flash | model id | Classifies section roles. |
REMEMBERSTACK_SUMMARY_MODEL | z-ai/glm-4.7-flash | model id | Section summaries. |
REMEMBERSTACK_E2_EXTRACT_MODEL | openai/gpt-5.6-luna | model id | Claim selection and extraction. |
REMEMBERSTACK_E3_NORMALIZE_MODEL | openai/gpt-5.6-luna | model id | Relation normalisation. |
REMEMBERSTACK_OBS_SMALL_MODEL | openai/gpt-5.6-luna | model id | Entity resolution. |
REMEMBERSTACK_FACT_MODEL | openai/gpt-5.6-luna | model id | Fact adjudication with the prompt engine. |
REMEMBERSTACK_FACT_ADJUDICATION_ENGINE | prompt | prompt or jev | Fact adjudication engine. Also read as REMEMBERSTACK_FACT_ENGINE. |
REMEMBERSTACK_FACT_CONFIDENCE_FLOOR | 0.75 | float 0–1 | Below this, both the new statement and the existing fact are kept. |
REMEMBERSTACK_FACT_FALLBACK_TO_PROMPT | false | bool | Fall back to prompt when a jev call fails. Also read as REMEMBERSTACK_TYPESAFE_FALLBACK_TO_PROMPT. |
REMEMBERSTACK_P1_EMBEDDING_MODEL | qwen/qwen3-embedding-8b | model id, 1,536-dim output | Every embedding, including search queries. Fixed once anything is embedded. |
Converters
See File formats and converters.
| Variable | Default | Type | Purpose |
|---|---|---|---|
REMEMBERSTACK_SELFHOST_CONVERSION_ROUTES | empty: the engine's format table (the table) | JSON object | MIME type → converter. Adds to the default table and overrides the entries it names. |
REMEMBERSTACK_MISTRAL_OCR_API_KEY | unset | secret | Mistral key; required by mistral_ocr and image_ocr_description. |
REMEMBERSTACK_MISTRAL_OCR_BASE_URL | https://api.mistral.ai | URL | Mistral API address. |
REMEMBERSTACK_MISTRAL_OCR_MODEL | mistral-ocr-latest | string | OCR model. |
REMEMBERSTACK_MISTRAL_OCR_TIMEOUT_S | 300 | float > 0 | OCR request timeout. |
REMEMBERSTACK_MISTRAL_OCR_MAX_DOCUMENT_BYTES | 50000000 | int > 0 | Largest file sent to OCR. |
REMEMBERSTACK_MISTRAL_OCR_INCLUDE_IMAGES | true | bool | Keep embedded images. |
REMEMBERSTACK_MISTRAL_OCR_TABLE_FORMAT | markdown | markdown or html | Table format. |
REMEMBERSTACK_MISTRAL_OCR_EXTRACT_HEADERS_AND_FOOTERS | true | bool | Extract headers and footers separately. |
REMEMBERSTACK_MISTRAL_OCR_CONFIDENCE_GRANULARITY | word | word or page | Confidence score detail. |
REMEMBERSTACK_MISTRAL_OCR_KEEP_PROVIDER_RESPONSE | true | bool | Keep the provider response as an artifact. |
REMEMBERSTACK_MISTRAL_OCR_PRICE_USD_PER_1000_PAGES | 1 | decimal ≥ 0 | Price used to record OCR cost. |
REMEMBERSTACK_IMAGE_DESCRIPTION_API_KEY | unset | secret | OpenRouter key for image descriptions; required by image_ocr_description. |
REMEMBERSTACK_IMAGE_DESCRIPTION_BASE_URL | https://openrouter.ai/api/v1 | URL | API address. |
REMEMBERSTACK_IMAGE_DESCRIPTION_MODEL | openai/gpt-6-luna | model id | Vision model; must accept image input. |
REMEMBERSTACK_IMAGE_DESCRIPTION_DETAIL | high | auto, low, high, original | Explicit provider observation resolution. High requests bounded preprocessing of the original image; exact sizing and supported values depend on the model. Changes refresh the description checkpoint while retaining matching OCR. |
REMEMBERSTACK_IMAGE_DESCRIPTION_REASONING_EFFORT | unset | reasoning effort | Optional vision effort; the LoCoMo profile uses low. |
REMEMBERSTACK_IMAGE_DESCRIPTION_PROVIDER_ONLY | unset | JSON string list | Approved endpoints in preference order. One entry disables fallback; multiple entries allow fallback only within the list. The LoCoMo profile uses ["openai/flex","openai"]. |
REMEMBERSTACK_IMAGE_DESCRIPTION_TIMEOUT_S | 120.0 | float > 0 | Request timeout. |
REMEMBERSTACK_IMAGE_DESCRIPTION_MAX_IMAGE_BYTES | 10000000 | int > 0 | Largest image. |
REMEMBERSTACK_IMAGE_DESCRIPTION_MAX_IMAGE_PIXELS | 40000000 | int > 0 | Largest width × height. |
REMEMBERSTACK_IMAGE_DESCRIPTION_MAX_IMAGE_WIDTH | 16000 | int > 0 | Largest width. |
REMEMBERSTACK_IMAGE_DESCRIPTION_MAX_IMAGE_HEIGHT | 16000 | int > 0 | Largest height. |
REMEMBERSTACK_IMAGE_DESCRIPTION_MAX_DESCRIPTION_CHARS | 16000 | int > 0 | Longest description. |
REMEMBERSTACK_IMAGE_DESCRIPTION_MAX_TOKENS | 4096 | int ≥ 1 | Output allowance of the description call. |
REMEMBERSTACK_IMAGE_DESCRIPTION_LANE_CONCURRENCY | 2 | int 1–2 | OCR and description in parallel (2) or in turn (1). |
Operations
| Variable | Default | Type | Purpose |
|---|---|---|---|
REMEMBERSTACK_SELFHOST_CORPUS_VERIFY_INTERVAL_S | 86400 | finite float > 0 | Cadence in seconds for digest-verifying corpus reconcile through the ordinary delayed-work ledger. |
REMEMBERSTACK_OPERATIONAL_SAMPLE_LIMIT | 20 | int 1–1000 | Cap on lists in remember ops inspect. |
REMEMBERSTACK_COST_EXPORT_BIND | unset | host:port, [ipv6]:port or unix:/path | Address of the HTTP cost export. Unset: off. |
REMEMBERSTACK_COST_EXPORT_TOKEN | unset | secret, ≥ 32 bytes | Bearer token for the cost export. |
Observability
| Variable | Default | Type | Purpose |
|---|---|---|---|
REMEMBERSTACK_SENTRY_DSN | unset | secret | Sentry-protocol DSN; unset keeps error tracking off. |
REMEMBERSTACK_SENTRY_ENVIRONMENT | deployment slug | string | Environment name on events. |
REMEMBERSTACK_SENTRY_SAMPLE_RATE | 1.0 | float 0–1 | Share of error events sent. |
REMEMBERSTACK_LANGFUSE_ENABLED, _HOST, _PUBLIC_KEY, _SECRET_KEY, _CA_FILE, _RUN_TAG | false / empty | Read only by the benchmark harness; no effect on the engine. | |
LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, LANGFUSE_HOST | empty | Read only by the benchmark harness; no effect on the engine. |
Knowledge plane and other unused settings
These are read by code in the engine that no shipped command runs. Setting them has no effect on a Compose deployment.
| Variables | Read by |
|---|---|
REMEMBERSTACK_K_PLANNER_PLANNER_MODEL, _PLANNER_MODEL_FAMILY, _REFLECTION_MODEL, _REFLECTION_MODEL_FAMILY, _TIMEOUT_SECONDS, _AUTO_APPLY_MAX_EXPECTED_IMPACT, _TRANSCRIPT_PREFIX | Knowledge-page planner |
REMEMBERSTACK_K_WRITER_MODEL, _TIMEOUT_SECONDS, _RESIDUE_CLAIM_LIMIT, _EVIDENCE_CLAIMS_PER_FACT, _TRANSCRIPT_PREFIX | Knowledge-page writer |
REMEMBERSTACK_K_DRIVER_MAX_PARALLEL_PAGES | Knowledge-page driver |
REMEMBERSTACK_K_CODEX_EXECUTABLE, REMEMBERSTACK_K_WRITER_CODEX_EXECUTABLE (default codex) | Knowledge-page agent adapter |
REMEMBERSTACK_SYNC_DEBOUNCE_QUIET_SECONDS (default 120), REMEMBERSTACK_SYNC_LANE (default steady) | Watched-directory sync loop |
REMEMBERSTACK_BACKFILL_BATCH_SIZE (default 500) | Backfill seeder |
REMEMBERSTACK_VERTEX_PROJECT_ID, _LOCATION (default global), _TIMEOUT_S, _MAX_COMPLETION_TOKENS (default 128000), _PRICE_TABLE_USD_PER_MILLION, _THROTTLE_RETRY_DELAYS_S | Google Vertex model adapter, used by the benchmark harness |