MemoDocumentation
EN

Proactive Learning & Calendar

Memo observes your patterns and conversation to become more helpful over time. The proactive system runs entirely on-device — no data leaves your machine for learning purposes.

Ambient Nudges (v3.3.3)

Beyond calendar/habit extraction, Memo now notices broader usage patterns and gently surfaces them — on by default at the subtle level.

  • A habit you state directly ("I code every night around 9") is trusted immediately, rather than needing to first show up statistically across many sessions like a passively-observed pattern does.
  • A nudge can appear woven naturally into one of Memo's own replies ("isn't it about your coding time?") instead of only as a separate suggestion banner. Whether it actually got raised, and whether you accepted or brushed it off, is judged by asking the model itself — not tied to any particular language's exact wording.
  • A real suggestion banner now exists on desktop (Yes / Not now / Stop asking) — previously there was no UI to see or respond to a pending suggestion.
  • Fully disabled under Incognito Mode, and under Minimal Mode unless you've specifically re-enabled it there (Settings → General now lets persona/system-prompt, capability disclosures, passive-feature disclosures, and proactive learning each be independently re-enabled, rather than all-or-nothing).

Routines — Scheduled Automations (v3.3.3)

Settings/sidebar → Routines lets you schedule something for Memo to do on its own, without having to ask each time.

  • Describe what you want and how often, in plain language — Memo turns it into a routine that fires in the background on schedule, either as a simple prompt or as a full tool-using agent run if enabled.
  • Works on both desktop and mobile — the mobile app delivers routine reminders as real, pre-scheduled local notifications, so they arrive even if the app isn't open.
  • Routines fire against your own device's timezone, captured at creation — that offset automatically resyncs on every reconnect, so travel or a DST change corrects itself instead of staying frozen.
  • Routine-generated text (system prompt for the run, "nothing due today" filler, notification titles) follows your actual language setting.

Self-Insight (/insight) (v3.3.3)

Ask Memo /insight directly, or let a Routine ask on a weekly schedule, and it looks back over your recent mood history and memory to describe any real pattern it's noticing — explicitly instructed not to invent one if there isn't enough to go on, in which case it just says so rather than guessing.

Two-Stage Intent Extraction

Every user message passes through a lightweight keyword pre-filter first. Only messages containing time-bound plans or habit language ("yarin saat 11", "her gun", "every day at 9") are forwarded to the LLM for full intent extraction. Routine chatter never triggers an LLM call.

Message → [Keyword Filter] → pass → [LLM Intent Extraction] → Event
                        → skip (no match) → ignored

Automatic Event Creation

When a plan is detected in conversation:

  • A calendar entry is created instantly
  • A notification confirms the event was added
  • You can delete it if the detection was wrong
  • Events include title, time, and a short summary

WhatsApp Integration

Intent extraction works on both sent and received WhatsApp messages. Memo learns from both directions of a conversation.

Observer Pattern Detection

Memo tracks when you use specific features and builds usage models. For example:

  • You always run a specific command after opening the agent
  • You check the calendar every morning at 9
  • You frequently ask for code reviews on Friday afternoons

These patterns are analyzed using circular statistics to detect temporal regularities (daily, weekly cycles).

Proactive Suggestions

Based on learned patterns, Memo can offer suggestions proactively. A notification appears with a suggested action — accept with one tap or dismiss.

Confidence Levels

Level Behavior
Off No suggestions
Low Only high-confidence patterns (90%+)
Medium Moderate patterns (70%+)
High All detected patterns, including low confidence

Settings

Configure in Settings → Learning:

Setting Description
enabled Master toggle for proactive learning
level off, low, medium, or high
Single Model Mode Route learning through one model instead of Orchestra

Privacy

Only summaries are stored — never raw message text. The keyword pre-filter ensures only relevant messages reach the LLM. All processing is local.