Configuration

August 5, 2026 · View on GitHub

Generated file. Do not edit by hand. Source: mnemosyne/core/config.py (ENV_VAR_MAP, DEFAULTS, REQUIRES_RESTART). Regenerate with python3 scripts/generate-docs.py.

Mnemosyne reads configuration from a YAML file and from environment variables. Precedence is config.yaml > environment variable > built-in default.

config.yaml lives at $MNEMOSYNE_DATA_DIR/config.yaml, falling back to $HERMES_HOME/mnemosyne/config.yaml and then ~/.hermes/mnemosyne/config.yaml. Nested YAML works, and leaf keys resolve too, so memory.mnemosyne.wm_max_items and wm_max_items are equivalent.

Keys marked restart are read once at startup. Changing them at runtime warns and does not take effect until the process restarts.

There are 106 configuration keys. A further 42 environment variables are read directly from the environment and are not settable in config.yaml; they are listed separately below.

Configuration keys (106)

Config key	Environment variable	Default	Restart	Description
`auto_migrate`	`MNEMOSYNE_AUTO_MIGRATE`	`true`	no	Run packaged migrations automatically on database open.
`auto_sleep_enabled`	`MNEMOSYNE_AUTO_SLEEP_ENABLED`	`true`	no	Run consolidation automatically on a background daemon.
`backup_dir`	`MNEMOSYNE_BACKUP_DIR`	(unset)	yes	Directory for `mnemosyne backups` snapshots.
`blob_dir`	`MNEMOSYNE_BLOB_DIR`	(unset)	yes	Directory for content-addressed blob storage.
`conflict_llm_api_key`	`MNEMOSYNE_CONFLICT_LLM_API_KEY`	(unset)	no	API key for the conflict detection endpoint.
`conflict_llm_base_url`	`MNEMOSYNE_CONFLICT_LLM_BASE_URL`	(unset)	no	Base URL for the conflict detection endpoint.
`conflict_llm_model`	`MNEMOSYNE_CONFLICT_LLM_MODEL`	(unset)	no	Model for conflict detection calls.
`cross_session`	`MNEMOSYNE_CROSS_SESSION`	`false`	no	Allow recall to return memories from other sessions.
`data_dir`	`MNEMOSYNE_DATA_DIR`	(unset)	yes	Root directory for all Mnemosyne data. Overrides the `$HERMES_HOME` default.
`db_path`	`MNEMOSYNE_DB_PATH`	(unset)	yes	Explicit path to the primary SQLite database file.
`default_owner`	`MNEMOSYNE_DEFAULT_OWNER`	(unset)	yes	Default owner ID for canonical facts and shared memory.
`default_scope`	`MNEMOSYNE_DEFAULT_SCOPE`	`session`	no	Default scope for new memories: `session` or `global`.
`degrade_batch`	`MNEMOSYNE_DEGRADE_BATCH`	`100`	no	Rows degraded per maintenance pass.
`embedding_api_key`	`MNEMOSYNE_EMBEDDING_API_KEY`	(unset)	no	API key for a remote embedding provider. Falls back to `OPENAI_API_KEY`.
`embedding_api_url`	`MNEMOSYNE_EMBEDDING_API_URL`	(unset)	yes	Base URL of an OpenAI-compatible embedding endpoint.
`embedding_dim`	`MNEMOSYNE_EMBEDDING_DIM`	`384`	yes	Override the embedding dimension. Must match the dimension stored in the existing vec tables.
`embedding_model`	`MNEMOSYNE_EMBEDDING_MODEL`	`BAAI/bge-small-en-v1.5`	yes	Embedding model name. Local fastembed model, or an API model when routed to a provider.
`embeddings_off`	`MNEMOSYNE_EMBEDDINGS_OFF`	`false`	no	Alias for `no_embeddings`.
`embeddings_via_api`	`MNEMOSYNE_EMBEDDINGS_VIA_API`	`false`	no	Force API mode for embeddings instead of inferring it from the model name.
`enhanced_recall`	`MNEMOSYNE_ENHANCED_RECALL`	`false`	no	Enable the enhanced recall pipeline with fact, graph, and episodic fusion.
`ep_limit`	`MNEMOSYNE_EP_LIMIT`	`50000`	no	Maximum rows retained in `episodic_memory`.
`fact_recall_enabled`	`MNEMOSYNE_FACT_RECALL_ENABLED`	`0` ¹ (declared `true`)	no	Enable structured fact matching during recall.
`fastembed_cache_dir`	`MNEMOSYNE_FASTEMBED_CACHE_DIR`	(unset)	yes	Cache directory for downloaded fastembed ONNX models.
`force_local`	`MNEMOSYNE_FORCE_LOCAL`	`false`	no	Skip the remote chain and use the local GGUF model directly.
`fts_weight`	`MNEMOSYNE_FTS_WEIGHT`	`0.3`	no	FTS5 rank weight in hybrid ranking.
`home`	`MNEMOSYNE_HOME`	(unset)	yes	Base Hermes home directory used to derive default paths.
`host_llm_enabled`	`MNEMOSYNE_HOST_LLM_ENABLED`	`false`	no	Route LLM work through a host-registered backend. See `core/llm_backends.py`.
`host_llm_model`	`MNEMOSYNE_HOST_LLM_MODEL`	(unset)	no	Model override passed to the host backend.
`host_llm_n_ctx`	`MNEMOSYNE_HOST_LLM_N_CTX`	`32000` ² (declared `2048`)	no	Context budget assumed for the host backend.
`host_llm_provider`	`MNEMOSYNE_HOST_LLM_PROVIDER`	(unset)	no	Provider override passed to the host backend.
`ignore_patterns`	`MNEMOSYNE_IGNORE_PATTERNS`	(unset)	no	Comma separated regexes; matching content is never stored.
`importance_weight`	`MNEMOSYNE_IMPORTANCE_WEIGHT`	`0.2`	no	Importance weight in hybrid ranking.
`lenient_fact_match`	`MNEMOSYNE_LENIENT_FACT_MATCH`	`false`	no	Match facts by substring instead of exact equality.
`llm_api_key`	`MNEMOSYNE_LLM_API_KEY`	(unset)	no	API key for the remote chat endpoint.
`llm_base_url`	`MNEMOSYNE_LLM_BASE_URL`	(unset)	no	Base URL of an OpenAI-compatible chat endpoint.
`llm_conflict_detection`	`MNEMOSYNE_LLM_CONFLICT_DETECTION`	`false`	no	Enable LLM-based contradiction detection during sleep.
`llm_enabled`	`MNEMOSYNE_LLM_ENABLED`	`true` ³ (declared `false`)	no	Enable LLM summarization during sleep consolidation.
`llm_fallback_api_key`	`MNEMOSYNE_LLM_FALLBACK_API_KEY`	(unset)	no	API key for the fallback chat endpoint.
`llm_fallback_base_url`	`MNEMOSYNE_LLM_FALLBACK_BASE_URL`	(unset)	no	Base URL for the fallback chat endpoint.
`llm_fallback_models`	`MNEMOSYNE_LLM_FALLBACK_MODELS`	(unset)	no	Comma separated models to try if the primary fails.
`llm_file`	`MNEMOSYNE_LLM_FILE`	(unset)	yes	GGUF filename within the repo.
`llm_max_tokens`	`MNEMOSYNE_LLM_MAX_TOKENS`	`2048` ⁴ (declared `512`)	no	Maximum output tokens per LLM summary.
`llm_model`	`MNEMOSYNE_LLM_MODEL`	(unset)	no	Model identifier for remote chat calls.
`llm_n_ctx`	`MNEMOSYNE_LLM_N_CTX`	`2048`	no	Context window for local GGUF inference.
`llm_n_threads`	`MNEMOSYNE_LLM_N_THREADS`	`4`	no	CPU threads for local GGUF inference.
`llm_repo`	`MNEMOSYNE_LLM_REPO`	(unset)	yes	Hugging Face repo holding the local GGUF model.
`llm_timeout`	`MNEMOSYNE_LLM_TIMEOUT`	`60`	no	Per-request timeout in seconds for LLM calls.
`no_embeddings`	`MNEMOSYNE_NO_EMBEDDINGS`	`false`	no	Disable dense retrieval entirely. Recall degrades to lexical FTS5.
`persona_daily_sync_hour`	`MNEMOSYNE_PERSONA_DAILY_SYNC_HOUR`	`3`	no	Local hour at which the daily persona sync runs.
`persona_enabled`	`MNEMOSYNE_PERSONA_ENABLED`	`true`	no	Inject L3 persona facts into the system prompt.
`persona_interval`	`MNEMOSYNE_PERSONA_INTERVAL`	`50` ⁵ (declared `10`)	no	Turns between persona refreshes.
`persona_token_cap`	`MNEMOSYNE_PERSONA_TOKEN_CAP`	`1500` ⁶ (declared `500`)	no	Token budget for injected persona facts.
`polyphonic_recall`	`MNEMOSYNE_POLYPHONIC_RECALL`	`false`	no	Enable the polyphonic multi-voice recall engine with RRF fusion.
`prefetch_content_chars`	`MNEMOSYNE_PREFETCH_CONTENT_CHARS`	`2000`	no	Truncate prefetched memory content to this many characters. `0` disables truncation.
`proactive_linking`	`MNEMOSYNE_PROACTIVE_LINKING`	`0` ⁷ (declared `true`)	no	Create cross-memory graph edges on insertion.
`query_intent`	`MNEMOSYNE_QUERY_INTENT`	`0` ⁸ (declared `true`)	no	Classify query intent and adjust scoring weights accordingly.
`recall_diagnostics`	`MNEMOSYNE_RECALL_DIAGNOSTICS`	`false`	no	Collect per-stage recall diagnostics. See `explain=True`.
`recall_extra_stopwords`	`MNEMOSYNE_RECALL_EXTRA_STOPWORDS`	(unset)	no	Extra comma or space separated stopwords for lexical recall.
`recency_halflife`	`MNEMOSYNE_RECENCY_HALFLIFE`	`168`	no	Hours until the recency decay factor halves.
`reflect_disabled_for_cron`	`MNEMOSYNE_REFLECT_DISABLED_FOR_CRON`	`true`	no	Skip reflection when running in a cron context.
`reflect_max_calls_per_session`	`MNEMOSYNE_REFLECT_MAX_CALLS_PER_SESSION`	`3`	no	Maximum reflection passes per session.
`shared_db_path`	`MNEMOSYNE_SHARED_DB_PATH`	(unset)	yes	Path to the shared cross-agent surface database.
`shmr_batch_size`	`MNEMOSYNE_SHMR_BATCH_SIZE`	`50`	no	Memories per SHMR harmonization batch.
`shmr_harmony_threshold`	`MNEMOSYNE_SHMR_HARMONY_THRESHOLD`	`0.60` ⁹ (declared `0.5`)	no	Harmony score required to merge a cluster into a belief.
`shmr_max_iterations`	`MNEMOSYNE_SHMR_MAX_ITERATIONS`	`3` ¹⁰ (declared `10`)	no	Maximum clustering iterations per SHMR batch.
`shmr_min_cluster_size`	`MNEMOSYNE_SHMR_MIN_CLUSTER_SIZE`	`2` ¹¹ (declared `3`)	no	Minimum memories required to form a cluster.
`shmr_model`	`MNEMOSYNE_SHMR_MODEL`	(unset)	no	LLM model for SHMR summarization. Empty uses the default chain.
`shmr_similarity_threshold`	`MNEMOSYNE_SHMR_SIMILARITY_THRESHOLD`	`0.7`	no	Cosine similarity required to cluster two memories.
`shmr_temperature`	`MNEMOSYNE_SHMR_TEMPERATURE`	`0.2` ¹² (declared `0.3`)	no	Sampling temperature for SHMR summarization.
`skip_contexts`	`MNEMOSYNE_SKIP_CONTEXTS`	`cron,flush,subagent,background,skill_loop`	no	Comma separated host context names that should not write memories.
`skip_embeddings`	`MNEMOSYNE_SKIP_EMBEDDINGS`	`false`	no	Alias for `no_embeddings`.
`sleep_batch`	`MNEMOSYNE_SLEEP_BATCH`	`5000`	no	Working memory rows processed per sleep cycle.
`sleep_model_refresh_auto_apply`	`MNEMOSYNE_SLEEP_MODEL_REFRESH_AUTO_APPLY`	`true`	no	Apply high-confidence proposals without review.
`sleep_model_refresh_auto_apply_min_confidence`	`MNEMOSYNE_SLEEP_MODEL_REFRESH_AUTO_APPLY_MIN_CONFIDENCE`	`0.90` ¹³ (declared `0.7`)	no	Minimum confidence to auto-apply a proposal.
`sleep_model_refresh_categories`	`MNEMOSYNE_SLEEP_MODEL_REFRESH_CATEGORIES`	`user,workflow,project`	no	Comma separated canonical categories eligible for refresh.
`sleep_model_refresh_conflict_min_confidence`	`MNEMOSYNE_SLEEP_MODEL_REFRESH_CONFLICT_MIN_CONFIDENCE`	`0.98` ¹⁴ (declared `0.8`)	no	Minimum confidence to auto-apply a proposal that contradicts a current fact.
`sleep_model_refresh_conflict_min_evidence`	`MNEMOSYNE_SLEEP_MODEL_REFRESH_CONFLICT_MIN_EVIDENCE`	`3` ¹⁵ (declared `5`)	no	Minimum supporting memories for a contradicting change.
`sleep_model_refresh_enabled`	`MNEMOSYNE_SLEEP_MODEL_REFRESH_ENABLED`	`true`	no	Let sleep propose updates to canonical facts.
`sleep_model_refresh_max_tokens`	`MNEMOSYNE_SLEEP_MODEL_REFRESH_MAX_TOKENS`	`2048` ¹⁶ (declared `1024`)	no	Token budget for a refresh proposal.
`sleep_model_refresh_min_evidence`	`MNEMOSYNE_SLEEP_MODEL_REFRESH_MIN_EVIDENCE`	`2` ¹⁷ (declared `3`)	no	Minimum supporting memories before proposing a change.
`sleep_model_refresh_temperature`	`MNEMOSYNE_SLEEP_MODEL_REFRESH_TEMPERATURE`	`0.1` ¹⁸ (declared `0.3`)	no	Sampling temperature for refresh proposals.
`sleep_prompt`	`MNEMOSYNE_SLEEP_PROMPT`	(unset)	no	Override the consolidation summarization prompt.
`smart_compress`	`MNEMOSYNE_SMART_COMPRESS`	`true`	no	Use LLM summarization rather than truncation when degrading a tier.
`sp_max`	`MNEMOSYNE_SP_MAX`	`1000`	no	Maximum scratchpad entries retained.
`sync_encrypt`	`MNEMOSYNE_SYNC_ENCRYPT`	`false`	no	Encrypt sync payloads client-side with XChaCha20-Poly1305.
`sync_host`	`MNEMOSYNE_SYNC_HOST`	`127.0.0.1`	yes	Bind address for `mnemosyne sync-serve`.
`sync_key`	`MNEMOSYNE_SYNC_KEY`	(unset)	no	Passphrase used to derive the client-side encryption key.
`sync_port`	`MNEMOSYNE_SYNC_PORT`	`8765`	yes	Bind port for `mnemosyne sync-serve`.
`sync_remote`	`MNEMOSYNE_SYNC_REMOTE`	(unset)	yes	Remote sync server URL.
`sync_roles`	`MNEMOSYNE_SYNC_ROLES`	`user`	no	Conversation roles synced into memory. Defaults to user turns only.
`sync_turn_assistant_limit`	`MNEMOSYNE_SYNC_TURN_ASSISTANT_LIMIT`	`10`	no	Maximum assistant turns captured per sync pass.
`sync_turn_user_limit`	`MNEMOSYNE_SYNC_TURN_USER_LIMIT`	`10`	no	Maximum user turns captured per sync pass.
`temporal_halflife_hours`	`MNEMOSYNE_TEMPORAL_HALFLIFE_HOURS`	`24` ¹⁹ (declared `168`)	no	Hours until the temporal recall boost decays by half.
`tier1_weight`	`MNEMOSYNE_TIER1_WEIGHT`	`1.0`	no	Recall score multiplier for tier 1 (fresh) episodic memories.
`tier2_days`	`MNEMOSYNE_TIER2_DAYS`	`30`	no	Age in days at which an episodic memory degrades to tier 2.
`tier2_weight`	`MNEMOSYNE_TIER2_WEIGHT`	`0.5`	no	Recall score multiplier for tier 2 memories.
`tier3_days`	`MNEMOSYNE_TIER3_DAYS`	`180`	no	Age in days at which an episodic memory degrades to tier 3.
`tier3_max_chars`	`MNEMOSYNE_TIER3_MAX_CHARS`	`300`	no	Character budget for tier 3 compressed content.
`tier3_weight`	`MNEMOSYNE_TIER3_WEIGHT`	`0.25`	no	Recall score multiplier for tier 3 memories.
`vec_type`	`MNEMOSYNE_VEC_TYPE`	`int8`	yes	sqlite-vec element type: `int8`, `float32`, or `bit`.
`vec_weight`	`MNEMOSYNE_VEC_WEIGHT`	`0.5`	no	Vector similarity weight in hybrid ranking. Normalized with the next two to sum to 1.0.
`wm_bump_cap_hours`	`MNEMOSYNE_WM_BUMP_CAP_HOURS`	`24`	no	Maximum window in which a recall can refresh a row's recency.
`wm_max_items`	`MNEMOSYNE_WM_MAX_ITEMS`	`10000`	no	Maximum rows retained in `working_memory` before eviction.
`wm_pinned_ids`	`MNEMOSYNE_WM_PINNED_IDS`	(unset)	no	Comma separated memory IDs that are never evicted.
`wm_ttl_hours`	`MNEMOSYNE_WM_TTL_HOURS`	`168`	no	Age at which a working memory row becomes eligible for consolidation or trim.
`write_classifier`	`MNEMOSYNE_WRITE_CLASSIFIER`	(unset)	no	Write classifier mode controlling which content is stored.

Keys whose effective default bypasses `config.py` (19)

For these keys a module-level constant reads the environment variable directly with its own fallback, so the value in DEFAULTS is not what the runtime uses and a config.yaml entry alone does not reach the module. Treat the environment variable as authoritative.

This list is derived by scanning the package for os.environ.get("MNEMOSYNE_...", default) and diffing against DEFAULTS, so it cannot fall out of date.

Environment-only variables (42)

These are read with os.environ at their point of use and bypass MnemosyneConfig entirely. They cannot be set in config.yaml.

Environment variable	Description
`MNEMOSYNE_AUTHOR_ID`	Identifier recorded as the author of new memories.
`MNEMOSYNE_AUTHOR_TYPE`	Author kind: user, assistant, system.
`MNEMOSYNE_AUTO_SLEEP_TIMEOUT`	Seconds the Hermes provider waits for a background sleep to finish.
`MNEMOSYNE_BANK`	Default memory bank for CLI operations.
`MNEMOSYNE_BEAM_MODE`	Enable BEAM mode extensions in the polyphonic engine.
`MNEMOSYNE_BEAM_OPTIMIZATIONS`	Enable BEAM benchmark optimizations.
`MNEMOSYNE_BINARY_BONUS`	A/B toggle for the binary vector bonus in episodic scoring.
`MNEMOSYNE_BUSY_TIMEOUT_MS`	SQLite `busy_timeout` in milliseconds.
`MNEMOSYNE_CHANNEL_ID`	Channel or session identifier for memory scoping.
`MNEMOSYNE_CONTEXT_INCLUDE_CONSOLIDATED`	Include already-consolidated rows in assembled context.
`MNEMOSYNE_CROSS_TIER_DEDUP`	A/B toggle for deduplication across working and episodic results.
`MNEMOSYNE_EMBEDDING_DOC_PREFIX`	Prefix prepended to documents before embedding. Applied verbatim.
`MNEMOSYNE_EMBEDDING_QUERY_PREFIX`	Prefix prepended to queries before embedding. Applied verbatim.
`MNEMOSYNE_EMBEDDING_THREADS`	Thread count for local ONNX embedding inference.
`MNEMOSYNE_EXTRACTION_MODEL`	Model used by the cloud extraction client.
`MNEMOSYNE_EXTRACTION_PROMPT`	Override the fact extraction prompt.
`MNEMOSYNE_FACT_BONUS`	A/B toggle for the fact match bonus in episodic scoring.
`MNEMOSYNE_GRAPH_BONUS`	A/B toggle for the graph traversal bonus in episodic scoring.
`MNEMOSYNE_HOST_LLM_TIMEOUT`	Timeout in seconds for host LLM backend calls.
`MNEMOSYNE_IMPORTED_WEIGHT`	Veracity multiplier for imported memories.
`MNEMOSYNE_INFERRED_WEIGHT`	Veracity multiplier for inferred memories.
`MNEMOSYNE_MCP_BANK`	Memory bank used by the MCP server.
`MNEMOSYNE_MCP_TOKEN`	Bearer token for MCP SSE auth. Required for any non-loopback bind.
`MNEMOSYNE_PERSONA_FILE`	Path to an external persona facts file.
`MNEMOSYNE_PREFETCH_MODEL_SLOT_LIMIT`	Maximum canonical slots prefetched per turn.
`MNEMOSYNE_PREFETCH_MODEL_SLOT_MIN_OVERLAP`	Minimum token overlap for a canonical slot to count as relevant.
`MNEMOSYNE_PREFETCH_PROFILE`	Prefetch profile name, for example `general` or `coding`.
`MNEMOSYNE_SESSION_END_TIMEOUT`	Seconds allowed for end-of-session memory writes.
`MNEMOSYNE_SHUTDOWN_DRAIN_TIMEOUT`	Seconds allowed to drain pending writes on shutdown.
`MNEMOSYNE_STATED_WEIGHT`	Veracity multiplier for directly stated memories.
`MNEMOSYNE_SYNC_KEY_SOURCE`	Where the sync passphrase comes from: `keyring` or `prompt`.
`MNEMOSYNE_SYNC_MODE`	Sync mode selector used by the Hermes sync adapter.
`MNEMOSYNE_SYNC_TOKEN`	Bearer token presented to the sync server.
`MNEMOSYNE_SYNC_TURN_SLOW_THRESHOLD`	Milliseconds after which a turn sync is logged as slow.
`MNEMOSYNE_TOOL_WEIGHT`	Veracity multiplier for tool-produced memories.
`MNEMOSYNE_UNKNOWN_WEIGHT`	Veracity multiplier for memories of unknown provenance.
`MNEMOSYNE_USE_CAVEMAN`	Use the AAAK keyword compression fallback for consolidation.
`MNEMOSYNE_VERACITY_MULTIPLIER`	A/B toggle for applying veracity multipliers at all.
`MNEMOSYNE_VOICE_FACT`	Set to `0` to disable the polyphonic fact voice.
`MNEMOSYNE_VOICE_GRAPH`	Set to `0` to disable the polyphonic graph voice.
`MNEMOSYNE_VOICE_TEMPORAL`	Set to `0` to disable the polyphonic temporal voice.
`MNEMOSYNE_VOICE_VECTOR`	Set to `0` to disable the polyphonic vector voice.

Recall scoring weights

vec_weight, fts_weight, and importance_weight are normalized to sum to 1.0 at query time. They are the highest-leverage tuning knobs in the system.

They resolve at request time as config.yaml > MNEMOSYNE_*_WEIGHT > defaults. A config reload applies to the next request, and enhanced-recall cache entries are isolated by the effective weight snapshot.

fact_recall_enabled -- effective default 0, set in mnemosyne/core/beam.py, not the true declared in config.py. ↩
host_llm_n_ctx -- effective default 32000, set in mnemosyne/core/local_llm.py, not the 2048 declared in config.py. ↩
llm_enabled -- effective default true, set in mnemosyne/core/local_llm.py, not the false declared in config.py. Note the direction: config.py declares this off while the module defaults it on. ↩
llm_max_tokens -- effective default 2048, set in mnemosyne/core/local_llm.py, not the 512 declared in config.py. ↩
persona_interval -- effective default 50, set in mnemosyne/core/persona.py, not the 10 declared in config.py. ↩
persona_token_cap -- effective default 1500, set in mnemosyne/core/persona.py, not the 500 declared in config.py. ↩
proactive_linking -- effective default 0, set in mnemosyne/core/beam.py, not the true declared in config.py. ↩
query_intent -- effective default 0, set in mnemosyne/core/beam.py, not the true declared in config.py. ↩
shmr_harmony_threshold -- effective default 0.60, set in mnemosyne/core/shmr.py, not the 0.5 declared in config.py. ↩
shmr_max_iterations -- effective default 3, set in mnemosyne/core/shmr.py, not the 10 declared in config.py. ↩
shmr_min_cluster_size -- effective default 2, set in mnemosyne/core/shmr.py, not the 3 declared in config.py. ↩
shmr_temperature -- effective default 0.2, set in mnemosyne/core/shmr.py, not the 0.3 declared in config.py. ↩
sleep_model_refresh_auto_apply_min_confidence -- effective default 0.90, set in mnemosyne/core/model_refresh.py, not the 0.7 declared in config.py. ↩
sleep_model_refresh_conflict_min_confidence -- effective default 0.98, set in mnemosyne/core/model_refresh.py, not the 0.8 declared in config.py. ↩
sleep_model_refresh_conflict_min_evidence -- effective default 3, set in mnemosyne/core/model_refresh.py, not the 5 declared in config.py. ↩
sleep_model_refresh_max_tokens -- effective default 2048, set in mnemosyne/core/model_refresh.py, not the 1024 declared in config.py. ↩
sleep_model_refresh_min_evidence -- effective default 2, set in mnemosyne/core/model_refresh.py, not the 3 declared in config.py. ↩
sleep_model_refresh_temperature -- effective default 0.1, set in mnemosyne/core/model_refresh.py, not the 0.3 declared in config.py. ↩
temporal_halflife_hours -- effective default 24, set in mnemosyne/core/beam.py, not the 168 declared in config.py. ↩

Configuration keys (106)

Keys whose effective default bypasses config.py (19)

Environment-only variables (42)

Recall scoring weights

Footnotes

Keys whose effective default bypasses `config.py` (19)