Adaptive Learning

How Armbrain automatically improves its search results and memory extraction based on your usage patterns.

What This Is

Over time, Armbrain learns which types of memories are most useful to you and which are noise. When you hide a memory that surfaced in search results, that is a signal. When certain memory types consistently appear in your meeting preps and you keep them, that is another signal.

Adaptive learning collects these signals and adjusts two things:

1. Retrieval weights - How search results are ranked. If you consistently find stakeholder memories more useful than generic facts, stakeholder memories get a boost in future searches.
2. Extraction sensitivity - How aggressively the ingestion pipeline extracts certain memory types. If you rarely use campaign outcome memories, extraction becomes more selective for that type.

These adjustments are gradual and conservative. You will not notice a dramatic shift overnight. Instead, search results and extractions slowly become better tuned to how you actually work.

How It Works (Behind the Scenes)

Armbrain watches for feedback signals:

When enough signals accumulate (at least 10 observations), Armbrain computes adjustments. Each adjustment is capped at 30% change to prevent overcorrection from a small number of signals.

Adapted parameters are stored in your personal mind preferences, not in any client mind. This means your adaptations follow you across all clients.

Running an Adaptation Cycle

Adaptations happen when you ask for them:

"Run an adaptation cycle"

Armbrain reviews your recent usage signals and computes any adjustments. You will see a summary:

Adaptation complete:
- Retrieval weight: stakeholder +0.08, campaign_outcome -0.05
- Extraction sensitivity: no changes (insufficient observations)
- Based on 23 signals since last adaptation

When to Run It

• Monthly is a good cadence for most users. You need enough usage between cycles to generate meaningful signals.
• After a burst of activity - If you just onboarded 3 clients and did heavy searching, there may be useful patterns to capture.
• After changing your workflow - If you started using a new set of tools or changed how you prep for meetings, run a cycle to recalibrate.

You do not need to run it frequently. The system requires at least 10 observations before making any adjustment, so running it daily will usually produce no changes.

Checking Current Adaptations

To see what adjustments are currently active:

"Show me my adaptation status"

This returns:

• Current retrieval weight adjustments (which memory types are boosted or dampened)
• Current extraction sensitivity adjustments
• How many observations have been collected since the last cycle
• When the last adaptation ran

Resetting Adaptations

If your results feel worse after an adaptation, or if your workflow has changed significantly:

"Reset my adaptations"

This clears all learned adjustments back to defaults. Your search results and extraction will behave as if no adaptations were ever applied.

After a reset, start fresh: use Armbrain normally for a few weeks, then run another adaptation cycle once you have built up new usage patterns.

What You Do Not Need to Do

• No configuration required. Adaptive learning works from your normal usage. Just use Armbrain and it collects signals automatically.
• No training step. There is no "teach Armbrain" mode. Your everyday actions (searching, hiding memories, prepping meetings) are the training data.
• No per-client tuning. Adaptations are stored in your personal mind and apply across all clients. This reflects your preferences as a CMO, not any specific client's needs.

Tips

• Hide memories that are not useful. This is the strongest signal Armbrain gets. If a search result is irrelevant, hiding it improves future results.
• Do not overthink it. Adaptive learning is designed to work in the background. Use Armbrain normally and let the system tune itself.
• Check adaptation status if results feel off. If searches are returning unexpected results, check whether an adaptation shifted weights in a direction you did not expect. Reset if needed.
• Give it time. Meaningful adaptation requires weeks of usage, not days. The conservative caps (30% maximum adjustment, 10-observation minimum) are intentional safeguards against premature optimization.