Brain and synthesis

Brain turns observations into agent knowledge; use synthesis settings to control when that knowledge updates.

You can

• Inspect observations and versioned agent knowledge.
• Choose shared fields or venture-specific fields.
• Adjust update timing, thresholds, and the synthesis model.
• Freeze updates while preserving current knowledge.
• Review freshness, usage, and costs.

KNOWLEDGE LIFECYCLE

Evidence becomes versioned working knowledge

1. 01

   Observe

   Workflows and people record useful evidence.

2. 02

   Filter

   Completion and importance gates protect quality.

3. 03

   Synthesize

   Eligible evidence is combined with the current Brain.

4. 04

   Version

   An immutable version becomes active for the agent.

What becomes Brain knowledge?

An observation records a learned fact, decision, outcome, blocker, or correction. It retains venture/run context, importance, creation time, and synthesis state.

Record	Purpose
Observation	Evidence awaiting or supporting synthesis
Brain version	Saved knowledge snapshot for one agent
Active Brain	Version currently supplied to that agent
Original source	Authoritative strategy, repository, resource, provider, or project record

Synthesis checks eligibility before including observations.

Observation source	Eligibility
Tracked Developer run	Completed or finished and reworked; reworked evidence carries a caveat
Recorded executive run	Completed
Incomplete, failed, or cancelled tracked run	Withheld
No matching tracked run	Not excluded by this run-status filter
Human action	Shared context; can trigger updates for existing agent brains

Reworked findings may have been superseded. Human context can trigger synthesis even when individual agents have not reached their thresholds.

Choose shared or custom Brain fields

Agent templates define role fields, labels, icons, instructions, and metrics. Without a profile, the first eligible synthesis generates one from the shared role definition.

Ventures inherit those field definitions. Personality, observations, metric values, and knowledge remain specific to each venture.

1. Edit Brain synthesis fields on an agent template to change shared defaults.
2. Open Brain → Fields in a venture to inspect global or custom fields.
3. Save a complete field list there to create a venture override.
4. Choose Use global fields to remove that override while preserving other venture settings.

Choice	Effect on future syntheses
Inherited fields	Follow global profile edits
Custom fields	Remain unchanged until edited or reset
Empty custom field list	Keep only shared Brain sections
Older saved profiles	Remain custom until reset

Existing Brain versions keep their original labels and icons.

When does synthesis run?

Each agent has an observation threshold balancing freshness against cost. Strategic agents generally use smaller batches; high-volume agents wait for more observations.

Use the threshold shown in Brain settings for the agent. Venture overrides can change it.

Situation	Expected behavior
Executive's first synthesis	One eligible observation can be enough, including after personality-only onboarding
Later synthesis	Normally waits for the minimum observation count
Brain cleared after earlier updates	Earlier updates still count; clearing Brain does not restore the one-observation starting threshold
Older Brain with new observations	Can refresh below the usual minimum
Scheduled selection	Selects agents that worked recently

Lower thresholds update knowledge sooner and use more synthesis. Higher thresholds reduce frequency but leave knowledge waiting longer.

Only sufficiently important, eligible observations are included. Unfinished or failed tracked runs do not silently become durable guidance.

Schedule, sync, or freeze updates

New venture schedules and newly created global fallback schedules default to every six hours. Saved cadences stay unchanged.

Each configured venture has a synthesis schedule. The cadence control offers frequent updates through weekly, plus Frozen.

Control	Result
Scheduled cadence	Batches eligible observations
Sync all now	Requests a manual update when available
Frozen	Pauses the venture schedule and excludes global fallback updates
Unfreeze	Resumes synthesis using accumulated eligible observations

Freezing disables Sync all now and pauses freshness aging at the freeze instant. Observations and Brain versions are retained.

Choose the synthesis model

Use Synthesis Model in Brain settings. General Developer or Content model settings do not change synthesis.

One provider/model pair serves all agents in a venture synthesis run.

Priority	Configuration
First	Venture synthesis provider and model
Second	Organization synthesis provider and model
Third	Platform synthesis default

Both provider and model must exist at the chosen level. Organ always takes both from the same settings level.

What happens during synthesis?

The run selects eligible venture/agent pairs and applies the chosen model.

For each agent, it considers:

• Existing Brain knowledge, when present.
• Eligible new agent observations.
• Shared eligible human observations.
• Agent configuration and importance thresholds.
• Available venture and repository context.

The result creates an immutable version, activates it, and links contributing observations. Token and cost records are saved alongside it.

Shared human observations finish processing after applicable agent syntheses complete. Failed or empty synthesis should leave observations awaiting processing.

Synthesis keeps role-specific knowledge compact and checks new technical claims against the repository. It preserves human constraints, rationale, ownership, and evidence for retiring temporary workarounds.

Where do synthesis costs appear?

Usage can include input, output, cache creation, cache reads, model, and estimated cost. Availability depends on the usage the provider reports.

Venture totals combine these calls using the same token prices and calculation as other work.

View	Includes
Brain version and agent analytics	Recorded synthesis cost
Venture model totals	Synthesis model usage
Month-to-date spend and spend-cap comparison	Synthesis cost
Organization token meter	Usage where configured

Execution compute is reported separately. See Compute, tokens, and model selection for formulas and billing boundaries.

Read Brain analytics

Measure	Available detail
Observation counts	Lifetime, seven-day, and 30-day totals per agent
Synthesis history	Count, last update, and cumulative cost
Activity	30-day observation and synthesis series
Recent evidence	High-importance observations
Version/outcome correlation	Available with sufficient venture age and matched run history

Correlation does not prove causation. Review important conclusions against their original sources.

Correct outdated knowledge

1. Correct the original strategy, resource, repository, or external record first.
2. Record a human observation when the correction should influence agent behavior.
3. Use Sync all now only when waiting for the saved cadence is costly.

Need	Consider
Noisy or repetitive observations	Raise thresholds
Strategic decisions need fresher context	Lower thresholds
Preserve stable knowledge during sensitive work	Freeze updates
Store credentials	Use encrypted resources rather than Brain

How do recurring jobs remember previous runs?

Recurring schedules can keep job-specific memory separately from the agent's general Brain. Later occurrences receive the purpose, lessons, open threads, and maintained workspace files.

Troubleshoot delayed or incorrect updates

If you see…	Why	What to do
Knowledge waiting for synthesis	Count, importance, or run eligibility requirements	Check agent thresholds and run completion
Update delayed	Model, provider, or execution failure	Inspect the synthesis run and last agent result
Empty or failed result	Synthesis did not produce usable knowledge	Inspect before repeating manual runs
Freshness stops aging	Venture is frozen	Unfreeze when updates should resume
Incorrect conclusion	Synthesis is not proof of accuracy	Correct the original source and record the correction

Brain is working knowledge. Strategy, credentials, repository state, and provider records remain authoritative in their original systems.