AtomicMemory offers durable, inspectable memory systems for AI applications. It extracts structured facts from conversations, detects contradictions during ingestion, and uses hybrid retrieval combining vector similarity with BM25 full-text search. The system tracks temporal claim versions through supersession operations. Deployment options include managed Hosted Cloud or self-hosted Core via Docker on local machines. Open-source under Apache-2.0 with TypeScript SDK and framework adapters available.
Should you self-host atomicmemory?
atomicmemory scores 54/100 on selfhost.directory (how scoring works) and ranks #35 of 125 in AI. It has 440 GitHub stars.
It shipped 5 stable releases in the past year. The latest, cli-v0.4.0, came out on 30 September 2026. New versions land roughly every 2 weeks.
Setup is moderate: expect some configuration beyond a single command. An official Docker image is available. It is Other-licensed and written mostly in Typescript.
ATOMICMEMORY_RUN_MIGRATIONS_ON_STARTUPDocker entrypoint only: set false when migrations run from a separate pre-deploy job. Defaults to true when omitted.
Docker entrypoint only: set false when migrations run from a separate pre-deploy job. Defaults to true when omitted.
AUDN_JSON_SCHEMAfalse keeps OpenAI-compatible requests on response_format=json_object.
false keeps OpenAI-compatible requests on response_format=json_object.
AUDN_MAX_TOKENS
CLOUD_JWKS_URLOptional: verify Cloud-issued RS256 JWTs for connected-local console access. JWKS_URL + ISSUER + AUDIENCE must be set together; when unset, only CORE_API_KEY is accepted.
Optional: verify Cloud-issued RS256 JWTs for connected-local console access. JWKS_URL + ISSUER + AUDIENCE must be set together; when unset, only CORE_API_KEY is accepted.
CLOUD_JWT_AUDIENCE
CLOUD_JWT_ISSUER
CLOUD_JWT_LEGACY_DEFAULT_MEMORY_USER_IDTemporary compatibility setting when the issuer omits memory_user_id:
Temporary compatibility setting when the issuer omits memory_user_id:
CLOUD_JWT_STATIC_KEY_FALLBACKOptional — when false (default), CORE_API_KEY is rejected once JWT verify is enabled.
Optional — when false (default), CORE_API_KEY is rejected once JWT verify is enabled.
CLOUD_PROJECT_IDOptional project binding. When set, tokens for other projects are rejected. When omitted (single-key local), Core trusts the token's own project_id.
Optional project binding. When set, tokens for other projects are rejected. When omitted (single-key local), Core trusts the token's own project_id.
CLOUD_TRACE_SYNC_BASE_RETRY_MS
CLOUD_TRACE_SYNC_BATCH_SIZE
CLOUD_TRACE_SYNC_CLAIM_STALE_MS
CLOUD_TRACE_SYNC_DEAD_LETTER_RETENTION_MS
CLOUD_TRACE_SYNC_ENABLED--- Optional Cloud trace sync (OSS Core → Cloud observability) --- Disabled when all vars are absent, or when CLOUD_TRACE_SYNC_ENABLED=false. Partial configuration (some vars set without ENABLED=true) fails startup.
--- Optional Cloud trace sync (OSS Core → Cloud observability) --- Disabled when all vars are absent, or when CLOUD_TRACE_SYNC_ENABLED=false. Partial configuration (some vars set without ENABLED=true) fails startup.
CLOUD_TRACE_SYNC_MAX_ATTEMPTS
CLOUD_TRACE_SYNC_MAX_PENDING
CLOUD_TRACE_SYNC_MAX_RETRY_MS
CLOUD_TRACE_SYNC_POLL_INTERVAL_MS
CLOUD_TRACE_SYNC_SENT_RETENTION_MS
CLOUD_TRACE_SYNC_SHUTDOWN_DRAIN_MS
CODEX_AUTH_PATHOptional: defaults to CODEX_HOME/auth.json or ~/.codex/auth.json.
Optional: defaults to CODEX_HOME/auth.json or ~/.codex/auth.json.
CORE_ADMIN_API_KEYOptional admin-only cleanup endpoint for disposable smoke/eval scopes. When both values are set, DELETE /v1/admin/scope accepts a JSON body `{ "user_id": "..." }` and deletes only matching test scopes. Use a different secret from CORE_API_KEY. Do not enable for general clients.
Optional admin-only cleanup endpoint for disposable smoke/eval scopes. When both values are set, DELETE /v1/admin/scope accepts a JSON body `{ "user_id": "..." }` and deletes only matching test scopes. Use a different secret from CORE_API_KEY. Do not enable for general clients.
CORE_API_KEY--- Core API auth --- Shared API key clients must send as `Authorization: Bearer <key>`. Rotate by restarting the server with a new value. Generate with: openssl rand -hex 32 Docker image local mode: auto-generated and persisted to $CORE_STATE_DIR/core-api-key (default /var/lib/atomicmemory/state/core-api-key) when omitted. Required in production/staging (RAW_STORAGE_DEPLOYMENT_ENV).
--- Core API auth --- Shared API key clients must send as `Authorization: Bearer <key>`. Rotate by restarting the server with a new value. Generate with: openssl rand -hex 32 Docker image local mode: auto-generated and persisted to $CORE_STATE_DIR/core-api-key (default /var/lib/atomicmemory/state/core-api-key) when omitted. Required in production/staging (RAW_STORAGE_DEPLOYMENT_ENV).
CORE_INSTANCE_ID
CORE_RUNTIME_CONFIG_MUTATION_ENABLED--- Runtime config mutation (dev/test only) --- Opt-in gate for PUT /memories/config. Leave unset in production — the route returns 410 Gone unless this is true. See https://docs.atomicstrata.ai/platform/consuming-core.
--- Runtime config mutation (dev/test only) --- Opt-in gate for PUT /memories/config. Leave unset in production — the route returns 410 Gone unless this is true. See https://docs.atomicstrata.ai/platform/consuming-core.
CORE_TEST_SCOPE_ALLOW_PATTERN
DATABASE_URL=postgresql://atomicmemory:atomicmemory@localhost:5433/atomicmemory--- Database --- Docker image users can omit DATABASE_URL, or set DATABASE_URL=embedded, to use the bundled single-container Postgres/pgvector instance. Source checkouts and production deployments should set a real Postgres connection string.
--- Database --- Docker image users can omit DATABASE_URL, or set DATABASE_URL=embedded, to use the bundled single-container Postgres/pgvector instance. Source checkouts and production deployments should set a real Postgres connection string.
EMBEDDING_API_KEY
EMBEDDING_API_URL
EMBEDDING_DIMENSIONS=1536Required — must match the actual output dimensions of your embedding model. mxbai-embed-large=1024, text-embedding-3-small=1536, Voyage 4=1024 Docker image local mode defaults this to `1536` when omitted.
Required — must match the actual output dimensions of your embedding model. mxbai-embed-large=1024, text-embedding-3-small=1536, Voyage 4=1024 Docker image local mode defaults this to `1536` when omitted.
EMBEDDING_MODEL
EMBEDDING_PROVIDER--- Embedding ---
--- Embedding ---
EXTRACTION_MAX_TOKENSDecode caps are whole safe positive integers (default 4096 / 2048; max 8192 / 4096). Not overridable via request config_override.
Decode caps are whole safe positive integers (default 4096 / 2048; max 8192 / 4096). Not overridable via request config_override.
EXTRACTION_PROMPT_VARIANT--- Extraction / AUDN (SLM latency) --- Default `full` keeps the complete extraction contract. `compact` is a reduced-capability opt-in for local SLM *extraction* only. It records user-stated facts only and omits: - assistant recommendations / advice - contact-info, named-entity, and date guarantees from the full prompt - short-input retention rules - headline, keywords, entities, relations, and i
--- Extraction / AUDN (SLM latency) --- Default `full` keeps the complete extraction contract. `compact` is a reduced-capability opt-in for local SLM *extraction* only. It records user-stated facts only and omits: - assistant recommendations / advice - contact-info, named-entity, and date guarantees from the full prompt - short-input retention rules - headline, keywords, entities, relations, and i
GOOGLE_API_KEY
GROQ_API_KEYRequired only for the matching provider selections:
Required only for the matching provider selections:
LISTEN_HOSTOptional bind address. Unset keeps Node's all-interfaces default (hosted/Docker). The macOS embedded launcher sets 127.0.0.1. Do not use HOST — it collides with the system hostname on many platforms.
Optional bind address. Unset keeps Node's all-interfaces default (hosted/Docker). The macOS embedded launcher sets 127.0.0.1. Do not use HOST — it collides with the system hostname on many platforms.
LITERAL_LIST_PROTECTION_ENABLED
LITERAL_LIST_PROTECTION_MAX_PROTECTED
LLM_API_KEY
LLM_API_URL
LLM_MODEL
LLM_PROVIDER--- LLM ---
--- LLM ---
MIGRATION_LOCK_TIMEOUT_MSDocker entrypoint only: forwarded to `migrate({ lockTimeoutMs })`. Increase for rolling deploys where another replica may hold the migration lock.
Docker entrypoint only: forwarded to `migrate({ lockTimeoutMs })`. Increase for rolling deploys where another replica may hold the migration lock.
OBSERVATION_DATE_EXTRACTION_ENABLED
OFFLINE_MODE--- Offline Personal profile (Radar C5) --- When OFFLINE_MODE=true, startup rejects any EMBEDDING_PROVIDER that makes external network calls (openai, voyage, openai-compatible). Use with EMBEDDING_PROVIDER=transformers or EMBEDDING_PROVIDER=ollama to guarantee zero external network calls. See docs/OFFLINE.md for the full profile.
--- Offline Personal profile (Radar C5) --- When OFFLINE_MODE=true, startup rejects any EMBEDDING_PROVIDER that makes external network calls (openai, voyage, openai-compatible). Use with EMBEDDING_PROVIDER=transformers or EMBEDDING_PROVIDER=ollama to guarantee zero external network calls. See docs/OFFLINE.md for the full profile.
OLLAMA_BASE_URL
OPENAI_API_KEY--- Provider credentials --- Required when either EMBEDDING_PROVIDER=openai or LLM_PROVIDER=openai.
--- Provider credentials --- Required when either EMBEDDING_PROVIDER=openai or LLM_PROVIDER=openai.
PORT=17350--- Server ---
--- Server ---
QUOTED_ENTITY_EXTRACTION_ENABLED
RAW_CONTENT_CODEC
RAW_CONTENT_CODEC_ACTIVE_KEY_ID
RAW_CONTENT_CODEC_KEYS
RAW_CONTENT_POLICY--- Ingest sensitivity policy (Radar C3) --- Server-side trust-boundary gate on /v1/memories/ingest{,/quick}. When `reject` (the default), core refuses to persist content marked `content_class: raw` AND content sent with no `content_class` at all (unstamped is treated as unknown/raw) — responding 422 raw_content_rejected. Single-user LOCAL deployments, where the store is not shared/hosted, may set
--- Ingest sensitivity policy (Radar C3) --- Server-side trust-boundary gate on /v1/memories/ingest{,/quick}. When `reject` (the default), core refuses to persist content marked `content_class: raw` AND content sent with no `content_class` at all (unstamped is treated as unknown/raw) — responding 422 raw_content_rejected. Single-user LOCAL deployments, where the store is not shared/hosted, may set
RAW_STORAGE_DEPLOYMENT_ENV=localRequired deployment posture for storage-policy gates. Use `local` for laptop/docker-compose development, `staging` for pre-prod, and `production` for hardened hosted deployments. Docker image local mode defaults this to `local` when omitted.
Required deployment posture for storage-policy gates. Use `local` for laptop/docker-compose development, `staging` for pre-prod, and `production` for hardened hosted deployments. Docker image local mode defaults this to `local` when omitted.
RAW_STORAGE_FILECOIN_DRIVER
RAW_STORAGE_FILECOIN_NETWORK
RAW_STORAGE_FILECOIN_PRIVATE_KEY
RAW_STORAGE_FILECOIN_SOURCE
RAW_STORAGE_FILECOIN_WITH_CDN
RAW_STORAGE_LOCAL_FS_ROOT
RAW_STORAGE_MODE--- Raw document storage (optional) --- Defaults to pointer-only mode. Use managed_blob only when configuring a concrete storage provider; misconfigured provider blocks fail at startup.
--- Raw document storage (optional) --- Defaults to pointer-only mode. Use managed_blob only when configuring a concrete storage provider; misconfigured provider blocks fail at startup.
RAW_STORAGE_PREFIX
RAW_STORAGE_PROVIDER
RAW_STORAGE_S3_ACCESS_KEY_ID
RAW_STORAGE_S3_BUCKET
RAW_STORAGE_S3_ENDPOINT
RAW_STORAGE_S3_REGION
RAW_STORAGE_S3_SECRET_ACCESS_KEY
SIMILARITY_THRESHOLD--- Internal retrieval tuning --- Defaults mirror the balanced adaptive policy. These are experimental knobs for benchmark sweeps and should not be treated as stable product config. Normalized fallback relevance floor for unscoped simple/medium non-temporal searches.
--- Internal retrieval tuning --- Defaults mirror the balanced adaptive policy. These are experimental knobs for benchmark sweeps and should not be treated as stable product config. Normalized fallback relevance floor for unscoped simple/medium non-temporal searches.
STORAGE_KEY_HMAC_SECRETHex-encoded HMAC secret used to derive PII-safe storage-key prefixes. Must be at least 64 hex chars (32 bytes of entropy). Rotating this invalidates existing managed-blob storage paths; only generate once per deployment. Generate with: openssl rand -hex 32 Docker image local mode provides a non-production default when omitted.
Hex-encoded HMAC secret used to derive PII-safe storage-key prefixes. Must be at least 64 hex chars (32 bytes of entropy). Rotating this invalidates existing managed-blob storage paths; only generate once per deployment. Generate with: openssl rand -hex 32 Docker image local mode provides a non-production default when omitted.
TEMPORAL_QUERY_CONSTRAINT_BOOST
TEMPORAL_QUERY_CONSTRAINT_ENABLED
TRUSTED_PROXY_MODETrusted-proxy identity guard (Radar C4). The shared CORE_API_KEY above authenticates the *caller process*, not the end user; `user_id` is asserted by the trusted caller (control plane / Radar daemon). When TRUSTED_PROXY_MODE=true, every user-scoped request carrying a `user_id` (body/query) or `X-AtomicMemory-User-Id` header MUST also carry the same value in the `X-AtomicMemory-Asserted-User` heade
Trusted-proxy identity guard (Radar C4). The shared CORE_API_KEY above authenticates the *caller process*, not the end user; `user_id` is asserted by the trusted caller (control plane / Radar daemon). When TRUSTED_PROXY_MODE=true, every user-scoped request carrying a `user_id` (body/query) or `X-AtomicMemory-User-Id` header MUST also carry the same value in the `X-AtomicMemory-Asserted-User` heade
audio.cppSelf-hosted C++ framework running 100+ audio models locally with CUDA, Vulkan, Metal, CPU backends, built-in WebUI, no Python deps.
AudioMuse-AIOpen-source, self-hosted sonic-analysis tool that builds playlists, a 2D music map and semantic search for Navidrome, Jellyfin, Emby, Plex, LMS and Lyrion libraries.
auliteSelf-hosted HTTP proxy that checks AI interactions for EU AI Act compliance, records them in a tamper-proof audit log, and produces legal-grade PDF reports.
Basic MemoryAI conversations that actually remember. Never re-explain your project to your AI again. Join our Discord: https://discord.gg/tyvKNccgqN
beamA self-hostable, open-source platform for running serverless AI workloads with GPU inference, code sandboxes, and background task queues.
beelzebubOpen-source, self-hosted deception runtime that turns SSH, HTTP, TCP, Telnet, MCP decoys into honeypots, configured via YAML rules, with LLM replies and Go plugins.
blindaiBlindAI is an open-source AI privacy solution using Intel SGX enclaves for confidential inference. Self-host models or query hosted ones via Python client.
BlinkoSelf-hosted AI note-taking app using RAG for intelligent search. Supports Markdown, native desktop apps on macOS/Windows/Android/Linux, PostgreSQL storage.
BokeBoxSelf-hosted private AI podcast studio turning videos, links, articles, and meetings into spoken episodes with customizable persona, voice, and MCP/plugin extensibility.
brain-dbSelf-hosted memory database for AI agents with fused semantic, lexical, graph, and temporal retrieval. Single Rust core, binary wire protocol, Apache-2.0.
brain.mdSelf-hosted markdown notes with MCP server for AI agents. Single Bun binary, local vector search via WASM ONNX, per-folder permissions, optional auth. AGPL.
Breeze TTS 2Self-hosted open-weight TTS model with voice cloning, design from descriptions, and voice direction. Streaming REST API and CLI for real-time synthesis on NVIDIA GPUs.
0 comments