Background intelligence

Most of what makes Mem useful happens in the background. Entity extraction, EVOLVES detection, contradiction flagging, Crystal synthesis, Working Memory updates: these run without you doing anything. You save a memory or import a thread, and the system takes it from there.

This page explains what runs, when it runs, and what keeps it from going off the rails.

Two kinds of triggers

Background tasks fall into two categories based on what starts them.

Scheduled tasks

These run on a clock, regardless of what you did today.

• Daily briefing runs early morning. Reviews recent activity, generates insights, flags contradictions, and writes a fresh Working Memory briefing for each active space. The Default space keeps the familiar ~/ai-now/memory.md compatibility file.
• Crystallization review runs weekly. Looks for clusters of related memories that could be synthesized into a Crystal.
• Insight detection runs weekly. Searches for cross-domain connections and patterns across your knowledge base.
• Skill suggestions run every 3 days by default. Looks for repeated ways of working that may be worth teaching your AI as a Skill.
• Rule suggestions run every 3 days by default. Looks for repeated preferences and standing rules your agents may need to follow. Suggestions stay as drafts until you review them.
• Memory compaction runs weekly when enabled. Reviews redundant memories and consolidates confirmed duplicates.
• Label consolidation runs weekly when enabled. Reviews duplicate labels, including labels written in different languages, before merging them.
• Community detection runs periodically. Rebuilds the entity graph's community structure for community-based search.
• Memory freshness runs daily when enabled. Recalculates freshness and confidence scores for all memories.
• Memory type review runs every 3 days by default in small batches. Helps file memories as facts, decisions, procedures, learnings, or events so agents can retrieve the right context.

Event-driven tasks

These fire in response to something you did, after a short delay.

• EVOLVES detection is triggered when you save a new memory. The system checks whether it replaces, enriches, confirms, or challenges something you already know. See knowledge evolution.
• Entity extraction is triggered alongside EVOLVES detection. Extracts entities and relationships for the knowledge graph.
• Memory type review is triggered in a small batch after new memories arrive. It improves how memories are organized without rewriting your saved memory text.
• Working Memory refresh is triggered when new memories arrive. Updates the current space's Working Memory so connected agents see new context quickly.
• Cluster evaluation is triggered after EVOLVES edges are created. Checks whether a new cluster has reached the threshold for Crystal formation.

The cascade

These are not independent. A single action can trigger a chain:

You save a memory → EVOLVES detection runs → finds a "confirms" link to an older memory → cluster evaluation fires → discovers three related memories now form a strong enough cluster → a Crystal is created.

Each step in the cascade has its own delay, so the system batches work rather than reacting to every individual event instantly. If you save five memories in quick succession, the system processes them together rather than running five separate analysis passes.

Guardrails

Background intelligence burns LLM tokens. Without limits, a burst of activity could drain your token budget or produce low-quality results from rushing through too many tasks. Four layers prevent this.

Debouncing. Event-driven tasks wait before executing. If another event of the same type arrives during the wait, the timer resets. This batches rapid-fire events (like importing a conversation with many messages) into a single analysis pass.

Rate limiting. A maximum number of LLM-powered tasks can run per hour. If the queue exceeds this, tasks are delayed rather than dropped.

Token budget. You can set hourly and daily token limits. When the budget runs out, the system pauses LLM tasks until the next period. Direct-function tasks (like decay refresh and community detection) are not affected because they do not use LLM calls.

Quality gates. The system suppresses low-value output. If a daily briefing produces zero insights, zero Crystals, and zero flags, it stays silent rather than generating a "nothing to report" card. Insight detection checks against the last two weeks of previous insights to avoid duplicates. Crystal formation requires three or more converging sources before creating anything.

Context injection

Each background task receives pre-computed context before it starts. The daily briefing, for instance, gets a digest of the past week's activity, yesterday's Working Memory, graph statistics, and recent resolution patterns. This avoids expensive LLM discovery steps and keeps each task focused.

Context is capped to prevent prompt bloat. If the pre-computed context exceeds the cap, lower-priority sections are trimmed first.

Working Memory

Working Memory is the daily briefing's most visible output. Every morning, the system archives yesterday's working memory and writes a fresh one based on recent activity.

The Default space keeps a file at ~/ai-now/memory.md. If you use spaces, other spaces get their own Working Memory briefings behind the same Mem APIs and integrations. This is how Claude Code, Grok Build, Cursor, Codex, and other tools know what you have been working on and what decisions you have made recently.

Working Memory also updates during the day. When you save new memories, the system refreshes the file (after a longer delay than other event-driven tasks, since it is more expensive to run). You can also edit it manually.

Processing settings

The Processing settings page gives you two controls for most tasks:

• Run starts that task once now.
• The switch controls whether Mem may run that task automatically in the future.

Small maintenance runs from Settings apply their safe, reviewed changes. For example, Memory Type Review only applies high-confidence type fixes from a bounded batch.

The green dot in Settings means a background task is currently running. Open Settings → Processing to see the current task name, when it started, and any available progress.

Recent runs show the latest task outcomes and reported or estimated token usage. Non-LLM housekeeping tasks show 0 tokens; LLM-backed tasks show approximate usage, and mark provider calls whose exact usage was not reported.

Usage boundaries

The AI usage card in Settings → Processing is a local safety ledger, not a universal bill.

It tracks:

• automatic Mem background jobs that run on this device
• built-in foreground AI sessions such as AI Now and Graph Intelligence, when the provider reports usage
• Plus managed-AI credit, when you are signed in to a Plus plan

It does not track every AI call made around Mem. Browser or Exchange agents, browser extensions, external IDE agents, and third-party hosts may call models through their own runtime or provider account. Those costs can be real, but they belong to that host's own bill or provider dashboard, not this local Mem ledger.

If you use your own API key, the provider's dashboard is still the source of truth for billing. Mem's local numbers help you cap automatic work and notice unusual activity, but they are not a replacement for the provider invoice.

Some providers do not return token usage for every response. When that happens, Mem marks the run as partially reported so you know the total may be lower than the real spend.

Not every task has the same risk profile:

• Safe housekeeping: Memory Freshness and Memory Type Review improve which memories surface first and how they are filed. They do not rewrite your saved memory text.
• Review-first maintenance: Memory Maintenance can prepare a Timeline review when old or overlapping memories may be adding noise. The refresh itself never archives, merges, deletes, or rewrites memories.
• Merging tasks: Memory Compaction and Label Consolidation can merge graph records after review. They are off by default; turn them on only if you want Mem to keep those graphs tidier automatically.
• Moving aside: Low-risk facts and events can be moved out of everyday recall only after review. Preferences, decisions, procedures, plans, learnings, rules, identities, and context memories are not mechanically archived; they are treated as material for semantic organization instead.

For normal use, leave Background Intelligence on and keep the defaults. If you run a server or VPS install, use the token limits on the same page to cap background AI cost.

Large maintenance sweeps, such as full memory-type reclassification, can use substantial AI credit and touch many memories. Run them from the CLI, preview the result first, then apply it only when you understand the changes:

nmem memories reclassify-types --target-type procedure --full-scan --wait
nmem memories reclassify-types --target-type procedure --full-scan --apply --wait

Task controls

Every background task has its own toggle. You can turn off EVOLVES detection while keeping the daily briefing, or disable insight detection while keeping entity extraction. The master toggle disables everything at once.

Token budgets and scheduling (briefing hour, community detection interval) are also configurable. Background Intelligence requires a remote LLM to be configured, since the tasks run on your machine and need a model to reason with.

Next steps

• Memory decay covers the decay refresh task and how scores are calculated
• Knowledge evolution explains EVOLVES detection in detail
• Crystals explains crystallization review and what triggers Crystal creation