.env.example

# Docker Compose bootstrap. Replace all three values before first use.
LORE_DB_ADMIN_PASSWORD=change-this-admin-password
LORE_DB_RUNTIME_PASSWORD=change-this-runtime-password
LORE_DB_MAINTENANCE_PASSWORD=change-this-maintenance-password
LORE_BIND_ADDRESS=127.0.0.1
LORE_PORT=3000

# Manual Node deployment: use the non-owner runtime login provisioned by
# `bun run db:bootstrap`. It must be a member of the NOLOGIN `lore_app` role.
# DATABASE_URL=postgres://lore_runtime:change-me@localhost:5432/lore
# DATABASE_POOL_SIZE=10
# LORE_MAINTENANCE_DATABASE_URL=postgres://lore_maintenance_runtime:change-me@localhost:5432/lore
# LORE_MAINTENANCE_POOL_SIZE=2
# LORE_MAINTENANCE_CONCURRENCY=1
# LORE_MAINTENANCE_POLL_MS=1000
# LORE_MAINTENANCE_SWEEP_MS=300000
# Old embedding generations remain rollback-capable for this bounded interval.
# LORE_EMBEDDING_ROLLBACK_SECONDS=604800

# `db:bootstrap` connects through DATABASE_URL as a migration owner and provisions
# separate request and maintenance logins. These values are not used by either
# runtime process after provisioning.
# LORE_RUNTIME_ROLE=lore_runtime
# LORE_RUNTIME_PASSWORD=change-me
# LORE_MAINTENANCE_ROLE=lore_maintenance_runtime
# LORE_MAINTENANCE_PASSWORD=change-me

# Optional deployment branding.
APP_TITLE=Memory for humans and agents
APP_SUBTITLE=Durable recall inside a database-enforced boundary.

# Auth fails closed. This pair is only for local, private development.
AUTH_MODE=none
ALLOW_INSECURE=1
LORE_LOCAL_SUBJECT=local
LORE_LOCAL_DISPLAY_NAME=Local User
LORE_LOCAL_EMAIL=local@example.com

# Self-host operators choose one deployment-wide provider and model. Lore v1 fixes
# the vector protocol at 1024 dimensions; there is no dimension setting.
LORE_EMBEDDING_PROVIDER=ollama
LORE_EMBEDDING_MODEL=qwen3-embedding:0.6b
LORE_EMBEDDING_TIMEOUT_MS=120000
# During a zero-downtime model migration, set this complete pair on the maintenance
# process. Keep the serving provider/model and both providers' credentials there;
# one sequential worker drains serving writes and the building generation.
# LORE_EMBEDDING_BUILD_PROVIDER=google
# LORE_EMBEDDING_BUILD_MODEL=gemini-embedding-2
# Cosine distance gate for dense candidates. Calibrate against a versioned suite.
LORE_SEMANTIC_DISTANCE_THRESHOLD=0.5
# Optional indexed exact-entity recall channel; disabled until benchmarked.
LORE_ENTITY_ALIAS_RECALL=0
# Optional adjacent chunks included in returned/reranked evidence (0-2).
LORE_EVIDENCE_NEIGHBOR_CHUNKS=0
# Highest-scoring chunks retained per visible Memory before adjacent expansion (1-5).
LORE_EVIDENCE_TOP_CHUNKS=1
# Benchmark-only first-stage depth diagnostic. It does not change deployment search
# limits and should be measured before paying to rerank a wider candidate pool.
# LORE_BENCHMARK_RETRIEVAL_LIMITS=10,20,50,100
# LoCoMo-only natural-boundary ablation. These fields come from the pinned source
# sessions; they are not inferred from answers and do not enable a production mode.
# LORE_BENCHMARK_CONTEXT_GROUP_KEY=sessionNumber
# LORE_BENCHMARK_CONTEXT_GROUP_ORDINAL_KEY=sessionTurn
# LORE_BENCHMARK_CONTEXT_GROUP_BASE_LIMIT=20
# LORE_BENCHMARK_CONTEXT_GROUP_MAX_GROUPS=3
# Optional iterative pseudo-relevance feedback depth for multi-hop retrieval (0-3).
# Each round uses the accumulated query chain and the strongest-overlap sentence
# from one newly retrieved RLS-visible passage. Novel results append behind the
# retained order and every round reruns the same filters. Keep 0 until benchmarked.
LORE_RETRIEVAL_FEEDBACK_QUERIES=0
# Optional temporal reciprocal-rank fusion across the second-stage candidate pool.
# 0 disables it; calibrate 0..1 on temporal/conflict-resolution evaluation before use.
LORE_RETRIEVAL_RECENCY_WEIGHT=0

# Ollama-only connection settings:
OLLAMA_BASE_URL=http://127.0.0.1:11434
# Docker Compose reaches the host through this separate address.
LORE_DOCKER_OLLAMA_BASE_URL=http://host.docker.internal:11434
OLLAMA_KEEP_ALIVE=0

# Optional privacy-safe OpenTelemetry export. Lore spans use bounded operation,
# outcome, and error-class attributes only; they never add Memory content, query,
# or tenant/resource identifiers. Leave unset to keep the SDK export disabled.
# OTEL_EXPORTER_OTLP_ENDPOINT=http://127.0.0.1:4318
# OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=http://127.0.0.1:4318/v1/traces
# OTEL_EXPORTER_OTLP_HEADERS=authorization=Bearer%20replace-me

# Operator-only backup/restore commands use explicit one-shot environment values:
# LORE_BACKUP_PATH=./backups/lore-YYYYMMDD.dump
# LORE_PITR_ARCHIVE_DIRECTORY=/secure/lore-wal-archive
# LORE_PITR_RESTORE_DRILL_CONFIRMED_AT=2026-08-01T12:00:00Z
# LORE_RESTORE_DATABASE_URL=postgres://lore_admin:...@localhost:5432/lore_restore
# LORE_RESTORE_CONFIRM=lore_restore

# Google Gemini recipe (replace the provider and model above):
# LORE_EMBEDDING_PROVIDER=google
# LORE_EMBEDDING_MODEL=gemini-embedding-2
# GEMINI_API_KEY=replace-with-a-server-side-key

# OpenAI recipe (replace the provider and model above):
# LORE_EMBEDDING_PROVIDER=openai
# LORE_EMBEDDING_MODEL=text-embedding-3-small
# OPENAI_API_KEY=replace-with-a-server-side-key

# Optional second-stage reranking. Supported providers: llamacpp, vllm,
# vllm-score, cohere, memos, or voyage.
# Lore retrieves under RLS first and sends only visible candidate passages.
# Rerank endpoints require HTTPS outside loopback or host.docker.internal.
# COHERE_API_KEY, MEMOS_API_KEY, and VOYAGE_API_KEY are provider-specific fallbacks
# for the common server-only LORE_RERANK_API_KEY. Failures fall back to fused order.
# RAM-conscious local recipe (llama.cpp serves the official ggml-org Q8 GGUF):
# brew install llama.cpp
# llama-server --hf-repo ggml-org/Qwen3-Reranker-0.6B-Q8_0-GGUF:Q8_0 --reranking --ctx-size 8192 --host 127.0.0.1 --port 8080 --no-webui --parallel 1 --n-gpu-layers all
# LORE_RERANK_PROVIDER=llamacpp
# LORE_RERANK_MODEL=ggml-org/Qwen3-Reranker-0.6B-Q8_0-GGUF:Q8_0
# LORE_RERANK_BASE_URL=http://127.0.0.1:8080
# LORE_DOCKER_RERANK_BASE_URL=http://host.docker.internal:8080
# LORE_RERANK_API_KEY=
# LORE_RERANK_TIMEOUT_MS=30000
# Candidate pool shared by provider reranking and optional temporal rank fusion.
# 20 is Lore's measured Qwen3-0.6B local starting point; recalibrate other providers.
# LORE_RERANK_CANDIDATE_LIMIT=20
# LORE_RERANK_MIN_SCORE=0.01
# LORE_RERANK_DIVERSITY_LAMBDA=0.85
# LORE_RERANK_WEIGHT=0.75
# llama.cpp uses the GGUF model's embedded rerank template; this override is vLLM/managed-only.
# LORE_RERANK_INSTRUCTION=Given a memory recall query, retrieve relevant memory passages that answer the query
# GPU/server alternative:
# LORE_RERANK_PROVIDER=vllm
# LORE_RERANK_MODEL=Qwen/Qwen3-Reranker-0.6B
# LORE_RERANK_BASE_URL=http://127.0.0.1:8000
# Apple Silicon vLLM-Metal score API alternative (serve with runner=pooling):
# LORE_RERANK_PROVIDER=vllm-score
# LORE_RERANK_MODEL=mku64/Qwen3-Reranker-0.6B-mlx-8Bit
# LORE_RERANK_BASE_URL=http://127.0.0.1:8000
# Experimental setwise scorer using one local Ollama chat model over the authorized
# candidate set. Benchmark before enabling; this is not the RAM-conscious default.
# LORE_RERANK_PROVIDER=ollama-listwise
# LORE_RERANK_MODEL=qwen3.5:4b
# LORE_RERANK_BASE_URL=http://127.0.0.1:11435
# LORE_RERANK_NUM_CTX=8192
# LORE_RERANK_MAX_OUTPUT_TOKENS=2048
# LORE_RERANK_MAX_DOCUMENT_CHARS=600
# LORE_RERANK_KEEP_ALIVE=5m
# Managed recipes:
# LORE_RERANK_PROVIDER=cohere
# LORE_RERANK_MODEL=rerank-v4.0-pro
# COHERE_API_KEY=replace-with-a-server-side-key
# LORE_RERANK_PROVIDER=memos
# LORE_RERANK_MODEL=memos-reranker-0.6b
# MEMOS_API_KEY=replace-with-a-server-side-key
# LORE_RERANK_PROVIDER=voyage
# LORE_RERANK_MODEL=rerank-2.5
# VOYAGE_API_KEY=replace-with-a-server-side-key

# Optional multi-query planning for counts, comparisons, temporal reasoning, and
# multi-hop recall. The planner sees only the question; the original query is always
# retained and every generated query is independently filtered through RLS.
# Supported providers: ollama, vllm, openai, or google. Native Ollama fixes
# thinking off, deterministic decoding, a 4K context, and a 256-token output cap.
# LORE_QUERY_PLANNER_PROVIDER=ollama
# LORE_QUERY_PLANNER_MODEL=qwen3.5:4b
# LORE_QUERY_PLANNER_BASE_URL=http://127.0.0.1:11434
# LORE_DOCKER_QUERY_PLANNER_BASE_URL=http://host.docker.internal:11434
# LORE_QUERY_PLANNER_KEEP_ALIVE=0
# LORE_QUERY_PLANNER_NUM_CTX=4096
# LORE_QUERY_PLANNER_API_KEY=
# LORE_QUERY_PLANNER_TIMEOUT_MS=30000
# Total retrieval-query budget including the original query (1-5).
# LORE_QUERY_PLANNER_MAX_QUERIES=3
# LORE_QUERY_PLANNER_INSTRUCTION=
# vLLM uses an OpenAI-compatible `/v1` base URL. OpenAI uses
# LORE_QUERY_PLANNER_API_KEY or the existing OPENAI_API_KEY. Google uses
# GEMINI_API_KEY by default. Leave the base URL unset for either managed API.

# Local benchmark-only fixed reader. This is not consumed by the Lore application.
# Supported providers: ollama, vllm, openai, or google. LongMemEval-V2 includes 29 question
# screenshots, so use a vision-capable model for a complete 451-question run.
# LORE_BENCHMARK_READER_PROVIDER=vllm
# LORE_BENCHMARK_READER_MODEL=your-fixed-reader-model
# LORE_BENCHMARK_READER_BASE_URL=http://127.0.0.1:8002/v1
# LORE_BENCHMARK_READER_API_KEY=
# LORE_BENCHMARK_READER_TIMEOUT_MS=120000
# LORE_BENCHMARK_READER_MAX_CONTEXT_CHARS=120000
# LORE_BENCHMARK_READER_MAX_OUTPUT_TOKENS=512
# LORE_BENCHMARK_READER_INSTRUCTION=
# Native Ollama only. Lore fixes the token context and keeps a warm model for at
# most five minutes during a run, then explicitly unloads it on normal exit.
# Set keep-alive to 0 for intentionally cold per-question latency measurements.
# LORE_BENCHMARK_READER_PROVIDER=ollama
# LORE_BENCHMARK_READER_MODEL=qwen3.5:4b
# LORE_BENCHMARK_READER_BASE_URL=http://127.0.0.1:11434
# LORE_BENCHMARK_READER_NUM_CTX=32768
# LORE_BENCHMARK_READER_THINKING=0
# LORE_BENCHMARK_READER_KEEP_ALIVE=5m
# Bounded in-process planner/reranker sweep cache (1-10000 entries).
# LORE_BENCHMARK_CACHE_ENTRIES=2000
# Add feedback variants to retrieval benchmark reports without changing production.
# LORE_BENCHMARK_RETRIEVAL_FEEDBACK_QUERIES=1
# LORE_BENCHMARK_RETRIEVAL_RECENCY_WEIGHT=0.6

# Optional benchmark-only LongMemEval-V2 judge. This scores the official
# abstention/gotcha cases and is independent from both Lore and the fixed reader.
# Supported providers: vllm, openai, or google.
# LORE_BENCHMARK_JUDGE_PROVIDER=vllm
# LORE_BENCHMARK_JUDGE_MODEL=your-evaluator-model
# LORE_BENCHMARK_JUDGE_BASE_URL=http://127.0.0.1:8002/v1
# LORE_BENCHMARK_JUDGE_API_KEY=
# LORE_BENCHMARK_JUDGE_REASONING_EFFORT=medium
# LORE_BENCHMARK_JUDGE_TIMEOUT_MS=43200000
# LORE_BENCHMARK_JUDGE_MAX_OUTPUT_TOKENS=4096
# MemoryAgentBench-only answer evidence depth (1-100).
# LORE_MEMORYAGENTBENCH_RETRIEVAL_LIMIT=10
# Benchmark-only structured candidate extraction + deterministic max(serial) policy.
# Single-hop runs once; multi-hop uses bounded CAR with a fresh RLS search per hop.
# LORE_MEMORYAGENTBENCH_CONFLICT_ASSEMBLY=0

# Production: remove ALLOW_INSECURE and choose one mode.
#
# Cloudflare Access / trusted proxy:
# AUTH_MODE=proxy
# ACCESS_TEAM_DOMAIN=yourteam.cloudflareaccess.com
# ACCESS_AUD=your-access-application-audience
#
# Temporary single-operator HTTP Basic protection. The Basic username is cosmetic;
# every accepted login maps to LORE_LOCAL_SUBJECT so it cannot select another User.
# AUTH_MODE=password
# UI_PASSWORD=change-me

# Cloudflare uses the HYPERDRIVE and MAINTENANCE_HYPERDRIVE bindings in
# wrangler.jsonc instead of database URLs. Both referenced configurations MUST be
# created with --caching-disabled and with their corresponding distinct login.