Review and Configure Memory

Numi keeps saved facts and an automatically updated profile for your account. Selected memory is included as background context in later model requests, across your chats. Review it when an answer relies on an outdated or incorrect fact.

Inspect and edit your memory

1. Open your profile menu at the bottom of the sidebar and select Settings.
2. Select Memory Profile to review Identity, Communication, Top of mind, Recent history, Earlier context, and Background.
3. Edit a section and select its checkmark (Save). Use its trash button to clear the section, then confirm.

The top of the panel shows the profile version, prompt token count, pending observations, last update, and memory model. History contains earlier versions. Input observations shown in history were considered by the pipeline; they are not necessarily facts retained in the profile.

Manual section edits save immediately and create a history entry. Consolidate now asks the background model to process the profile and pending observations; it does not finish synchronously. Reopen the panel after the job completes to check the result.

Use Saved Preferences in Settings to review or delete explicit facts stored by remember_fact. For instructions you write yourself, use Settings → Instructions.

Clear stored memory

In Settings → Memory Profile, select Clear all memory and read the confirmation before continuing.

This deletes your profile, saved facts, pending observations, chat summaries, and memory history. Chats and custom instructions remain. Messages from before the reset are excluded from subsequent memory extraction.

Clearing is not an opt-out: new messages can build memory again. The current UI has no per-user switch to disable automatic memory. An operator can disable the background pipeline for the instance as described below.

Operator: choose the memory model

Automatic extraction and consolidation are enabled by default and run in the background after completed conversation turns. They use a separate model from the model selected in a chat.

On the admin Models page, find Memory extraction and summaries, choose an available model, and select Save override. Use config removes the override and returns to MEMORY_SUMMARY_MODEL. The selected catalog model needs working credentials for its provider; the default uses OpenRouter.

Operator: tune or disable the pipeline

Set environment variables in the deployment configuration and restart the server. The main controls are:

Variable	Default	Effect
MEMORY_ENABLED	true	Enables automatic extraction and consolidation
MEMORY_SUMMARY_MODEL	deepseek/deepseek-v4.1-flash	Model when there is no admin override
MEMORY_EXTRACTION_GATE_ENABLED	false	Ask Jev to skip turns that are confidently not worth memorizing
MEMORY_EXTRACTION_GATE_SKIP_BELOW	0.15	Skip extraction below this memory-candidate probability
MEMORY_EXTRACTION_TRIGGER_EVERY_N_TURNS	2	Completed pending turns needed to trigger extraction
MEMORY_EXTRACTION_CONSOLIDATION_THRESHOLD	5	Pending observations needed to trigger consolidation
MEMORY_EXTRACTION_IDLE_FLUSH_MINUTES	60	Age at which idle turns and small pending queues become eligible
MEMORY_EXTRACTION_IDLE_FLUSH_MAX_AGE_HOURS	0	Oldest idle chats to process; 0 means no age cutoff
MEMORY_CONSOLIDATION_RETRY_COOLDOWN_MINUTES	60	Delay before retrying failed background work
MEMORY_USER_PROFILE_TOKEN_BUDGET	1500	Budget for profile facts included in a prompt

The idle worker checks every five minutes and processes bounded batches, so an eligible chat may not update immediately. Deferred observations are kept for review instead of being retried as ordinary pending work.

The optional extraction gate sends the user message to OpenRouter before the configured memory model. It fails open: uncertain answers, invalid responses, and request failures continue through the normal extractor. A skipped turn does not receive a generated per-chat summary.

To stop background extraction and consolidation, set MEMORY_ENABLED=false. This does not delete existing memory or stop stored context from being included in prompts. Explicit remember_fact and forget_fact tools are controlled separately by the tool allowlist and denylist.

See Adaptive memory settings for retention windows and the complete configuration reference.

Storage and API access

The profile and explicit facts are scoped to the signed-in user. Extraction processes completed turns; observations wait for consolidation before becoming profile facts. Only selected facts within the prompt budget are sent as stored context, rather than the entire history of observations and snapshots.

Use the authenticated memory endpoints to inspect or edit profiles, retrieve history and observations, clear memory, or request consolidation from a custom client. Memory included in a prompt is sent to the provider handling that request, so remove information you do not want used in future requests.