KMS knowledge aggregation (periodic AI-curated session summary notes)
A background pipeline that turns a period of session activity into curated notes in your vault: it scans recent sessions, picks the notable ones, pulls the key insights out with a cheap model, and writes one summary note per project into an Agent Insights folder. Opt-in and cost-capped.
On this page
What it is
Omniscio can automatically collect notable outputs from your coding sessions, run selective AI deep-dives on the most interesting ones, and synthesize everything into periodic summary notes (daily and/or weekly) written as markdown files inside your KMS vault's Agent Insights/ folder.
Knowledge aggregation is a background pipeline that turns your accumulated session activity into curated, human-readable knowledge notes. Instead of manually reviewing dozens of sessions, Omniscio's aggregation worker scans sessions from a time period, identifies the most notable ones (by engagement, cost, keywords like "fix", "bug", "refactor", "decision"), extracts key insights using a small, low-cost AI model, and writes a single summary note per project.
The feature is opt-in and cost-capped. All AI calls use a low-cost model, and a configurable micro-dollar cost cap prevents runaway spend. You do not need an Anthropic (Claude) API key of your own: the digest is paid the same way as session titles and the other built-in AI helpers — your own AI keys first, then your Omniscio credits. The periodic worker runs alongside the KMS service; it starts when KMS starts and stops when KMS stops.
Switching it on takes effect at the next app launch. If your Vault is already running when you flip AI Knowledge Digest on, no digest is produced — not even a catch-up for a day or week already missed — until you restart Omniscio, because the schedule is created when the Vault starts. Switching it off works straight away.
Everything here is gated by two settings: the master nothariEnabled setting (Settings → Features → Enable Vault) AND nothariAggregationEnabled (the aggregation-specific toggle). Both must be on for the worker to run.
Where to find it
Its output lands where you already read notes — an Agent Insights folder inside your KMS vault, one summary note per project. The pipeline itself is switched on in Settings, where it is opt-in and carries a spend cap you set.
How it behaves
Settings
Turn it on at Settings → Features → The Vault → AI Knowledge Digest. The rest of the card appears once that switch is on:
| On screen | What it does | Default and range |
|---|---|---|
| Daily digest | Writes one digest a day. | Off |
| Run daily at | The hour the daily digest runs (shown once Daily digest is on). | 6:00 AM, 0:00–23:00 |
| Weekly digest | Writes one digest a week. | On |
| Run weekly on | The day and hour the weekly digest runs (shown once Weekly digest is on). | Monday at 6:00 AM |
| Per-run cost cap | The most one run may spend, in dollars. | $0.10, $0.01–$10.00 |
| Deep-dive session limit | How many of the most notable sessions get the expensive transcript pass, per run. | 10, 1–50 |
| Digest sections | Which sections appear in the note: Key decisions, Patterns and themes, Bugs and fixes, Lessons learned, Open questions, Session activity summary, Vault activity. | All on except Open questions |
The hour fields are your local time. For agents and scripts, the same settings persist as nothariAggregationEnabled, nothariAggregationDaily, nothariAggregationDailyHour, nothariAggregationWeekly, nothariAggregationWeeklyDay (0 = Sunday), nothariAggregationWeeklyHour, nothariAggregationCostCapMicroUsd (integer micro-USD), nothariAggregationDeepDiveMax and nothariAggregationSections — the nothari prefix is KMS's internal persistence name.
All settings use the nothari prefix (the internal persistence name for KMS).
| Setting | Default | What it controls |
|---|---|---|
nothariAggregationEnabled |
false |
Master on/off for the aggregation feature |
nothariAggregationDaily |
false |
Run daily aggregation |
nothariAggregationWeekly |
true |
Run weekly aggregation |
nothariAggregationDailyHour |
6 |
Hour (0-23) to run daily aggregation |
nothariAggregationWeeklyDay |
1 |
Day of week (0=Sun, 1=Mon) for weekly runs |
nothariAggregationWeeklyHour |
6 |
Hour (0-23) to run weekly aggregation |
nothariAggregationTimeZone |
null (this computer's) |
Zone the schedule resolves against |
nothariAggregationCostCapMicroUsd |
100000 |
Max micro-USD per run ($0.10 default) |
nothariAggregationDeepDiveMax |
10 |
Max sessions to deep-dive per project |
nothariAggregationSections |
all on except openQuestions |
Which sections to include in notes |
What it writes
Each run writes one note per project into that project's own folder under Agent Insights/ — Agent Insights/{Project Name}/2026-07-12 Daily.md for a daily run, Agent Insights/{Project Name}/2026-W28 Weekly.md for a weekly one. When the period covers more than one project, a combined note for all of them is written to Agent Insights/Everything/.
A note can hold these sections, chosen in Settings: decisions, patterns, bugs, lessons, activity summary and vault activity — plus open questions, which is off by default. Its front matter records the period type, the date range, the session count and what the run cost.
A run that targets a note that already exists replaces that note.
Cost and safety
- All AI calls use a low-cost model (OpenRouter's
openai/gpt-5.6-luna) - What leaves the device: for each session it deep-dives (up to
nothariAggregationDeepDiveMaxper project), up to about 50,000 characters of transcript go to OpenRouter; the rest of the sessions in the period contribute their name and AI summary, and the synthesis step also receives the titles of the vault notes the period touched. AI-coaching and journal sessions are excluded — they are private projects, so they never enter extraction, synthesis or a digest note. - Who pays: the same order every built-in AI helper uses — your own AI keys first, then your Omniscio credits. No Anthropic API key of your own is required; if you have one, the spend is recorded against that account. When nothing can pay, the run fails with the standard "AI helpers need a key" notice rather than writing an empty note.
- After the app restarts, the first run can come right away: the worker catches up on a daily or weekly run it missed while the app was closed.
- A cost cap (
nothariAggregationCostCapMicroUsd) stops the pipeline mid-run if cumulative cost exceeds the limit - The cost cap cannot be bypassed from either door:
POST /kms/aggregation/runand thekms:aggregation-run-nowIPC channel both accept onlyperiodType, and both schemas are strict — a body carryingbypassCostCap(or any other extra key) is rejected, not quietly ignored. The Run now button on the digest card (Settings → Features → KMS) does not bypass it either: it asks for confirmation first, then runs and stops at the same per-run cap. - Cost is tracked per-run in the
nothari_aggregation_runsdatabase table using integer micro-USD (the monetary-integer-twin pattern). The figure is what the model calls actually billed: when one call is retried on the fallback model, both attempts are charged, so both count toward the cap and into the recorded run cost. - Every project is read, AI-coaching and journal sessions included. Unlike the daily digest, which leaves those sessions out, the digest studies them like any other project; the switch in Settings says so, together with the vendor (OpenRouter) and the 50,000-character transcript cap.
- What is sent to the model: for each deep-dived session, the last 200 messages of that session's own transcript are sent to OpenRouter's
openai/gpt-5.6-luna— system rows are dropped and each session's transcript is cut at 50,000 characters. Session transcripts contain whatever was written in them: anything you pasted into a session, file contents an agent printed, and tool output. Nothing else from your computer is sent, and no note is uploaded. - A failed period gets one paid attempt per app run. The worker does not re-try it on the next scheduled run in the same run of the app; it tries that period once more the next time the app starts, and only while that period is still the most recent completed one. If the app is restarted after a whole day or week has gone by, that period is not produced at all — trigger it from the CLI (or a manual run) if you want it.
- Runs that find no sessions in the period are recorded as
skipped_no_sessions(no AI cost) - Errors are recorded with their message for debugging
- The worker raises inbox alerts when the cost cap is hit or a run fails entirely (dedup keys
kms-aggregation-cost-capandkms-aggregation-error), each linking to KMS settings
Related
The vault these notes are written into, and the editor you read them in, is described on the KMS page. The per-note AI summaries are a separate feature on KMS summaries.
Last verified 2026-10-20