About Services Tools Blog Contact us

Always-On Local Agent

A local assistant that takes in notes on a timer and spends from a shared energy budget. It reuses a past answer before it asks Ollama.

  • Agentic Workflows
  • Cross-sector
  • Scheduled worker
  • Product
  • Local/private
  • 2026

The problem

An assistant that only replies when prodded forgets context and has no sense of cost. Asking Ollama about every incoming note is wasteful. A fix that worked once was not saved, so the next identical request paid for the same work again.

What it does

Text is saved as a note on the fastest timer, indexed by meaning, and earns an energy credit. The assistant tries cheaper steps first, which are repeating a past action or running a saved skill. A larger model runs only when those miss, and a useful answer can be saved as a new skill. When energy is low it skips the deeper thinking.

Outcomes

  • Repeated requests skip Ollama.
  • An energy meter shows the budget, and work is refused when it is low.
  • An answer that matched no existing skill can be saved as a new one.
  • Each timer writes a heartbeat row, so a stalled timer is easy to see.

Simplified mockup

Interactive preview with sample data. Nothing is sent to a server.

Studio desk

Awake idle
Energy
72
Urgency
Low
Curiosity
Steady
liveBeatEvery 4 seconds
idleTideEvery 30 seconds
idleFocusEvery 3 minutes
asleepSweepEvery hour
Model calls 0
Cheap path
A familiar note should not spend the model
  1. Embed the note
  2. Search stored habits
  3. Replay the past answer
Thinking path
Used only when the cheaper layers miss
  1. Check habits
  2. Check stored skills
  3. Call the local model
  4. Save a new skill
Reply feed
Remind me about the stair rod quote.
The stair rod quote is with North Quay Interiors, filed under the hallway fitting.

Four rhythms

locks held
Heartbeats
RhythmEveryLockLast beatWork
Beat4 secondsheld2 seconds agoStore the note and credit energy
Tide30 secondsfree12 seconds agoSpread activation and decay stale items
Focus3 minutesfree1 minute agoHabit, skill, then the local model
Sweep1 hourfree18 minutes agoPrune weak links and archive cold notes
Failures none

No rhythm has missed its lock. A stalled beat would show here instead of a silent gap.

Low energy

When the meter is low, the focus rhythm skips. Beat, tide and sweep keep their cheaper work.

Linked notes

Similarity
habit Stair rod quote Hallway fitting, North Quay Interiors
linked Fitting date Held until the linen arrives
pending New skill Saved when the model answers
Cold archive
NoteLink strengthState
Stair rod quote0.94warm
Fitting date0.71linked
Last year paint schedule0.12cold

How it works

  1. Send a sentence to the assistant and save it as a note.
  2. Run four timers, from a few seconds up to an hour, each with its own lock.
  3. Link similar notes, strengthen related ones and fade stale ones.
  4. Repeat a past action when the new note is very similar, or run a saved skill when keywords and meaning agree.
  5. Ask Ollama only if the cheaper steps do not answer, then spend from the energy budget.
  6. On the hourly pass, remove weak links and archive old notes.

Pipeline

flowchart LR
  A[Note] --> B[Embed]
  B --> C[Habit replay]
  C --> D[Stored skill]
  D --> E[Ollama]
  E --> F[New skill]
  G[Energy budget] --> E