Perception Layer¶

Quick Answer: Perception records structured content outcomes, groups them by dimensions, and produces a bounded summary for later decisions. It is an analytics and feedback layer, not a guarantee that an agent's content is correct or that every session has full history injected.

A bounded feedback loop¶

Record an outcome, generate insights after enough samples, and request a summary only when the next decision needs it. Keep the summary bounded so it informs the agent without overwhelming the task context.

Let your agent learn from content operation results — what worked, what didn't, and why.

The perception loop closes the gap between "agent does something" and "agent knows the result." It records structured outcome snapshots, generates statistical insights by dimension, and injects a perception summary into the agent's context so future decisions are data-informed.

How It Works¶

graph LR
    A[Content Published] --> B[Outcome Recorded]
    B --> C[Insights Generated]
    C --> D[Perception Injected]
    D --> E[Agent Sees History]
    E --> A

Record — after publishing content, record metrics (likes, saves, comments, views) and context (topic, format, publish time)
Analyze — group outcomes by dimension, compute statistics, determine confidence levels
Inject — build a perception summary that the agent sees in its context on next run

Quick Start¶

Record an Outcome¶

aios perception record \
  --content-id "note_abc123" \
  --platform xiaohongshu \
  --content-type note \
  --title "10 AI Tools for Productivity" \
  --metrics '{"likes":150,"comments":23,"saves":45,"views":2000}' \
  --context '{"topic":"AI工具","format":"图文","publishHour":20}'

Generate Insights¶

After recording several outcomes (minimum 3 per dimension group):

aios perception insights --min-sample 3

Output:

Generated 3 insights from 5 outcomes.
  [insight] topic=AI工具 avgLikes=123 avgSaves=53 confidence=low sampleSize=3
  [insight] format=图文 avgLikes=157 avgSaves=45 confidence=low sampleSize=4
  [insight] publishHour=20 avgLikes=123 avgSaves=53 confidence=low sampleSize=3

View Perception Summary¶

aios perception summary

# JSON output for programmatic use
aios perception summary --format json

Perception Summary Format¶

When the agent starts a new session, it sees a summary like this:

## Perception Layer

### Performance Summary
- Content Published: 12
- Avg Engagement: 238
- Trend: improving (+18%)
- Best Recent: "10 AI Tools" (engagement=380)

### Recent Outcomes
- [24h] "10 AI Tools" likes=150 saves=45 comments=23
- [24h] "Daily Life" likes=200 saves=30 comments=45

### Active Insights
- [topic:AI工具] avgLikes=150, confidence=high (n=8)
- [format:图文] avgSaves=45, confidence=medium (n=5)
- [publishHour:20] avgEngagement=280, confidence=low (n=3)

### Strategy Recommendations
- Focus on topic "AI工具" (highest engagement)
- Prefer format "图文" (highest save rate)
- Publish around 20:00 (best time slot)

Dimensions¶

Outcomes are grouped by these context dimensions:

Dimension	Description	Example
`topic`	Content topic/category	"AI工具", "恋爱", "情绪"
`format`	Content format	"图文", "视频", "vlog"
`publishHour`	Hour of publication (0-23)	20
`publishDayOfWeek`	Day of week	"Monday"
`contentType`	Content type	"note", "video"
`coverStyle`	Cover image style	"minimal", "illustration"

Confidence Levels¶

Insight confidence is based on sample size:

Level	Sample Size	Meaning
high	n >= 8	Reliable signal
medium	n >= 5	Promising pattern
low	n >= 3	Early indication
insufficient	n < 3	Not enough data

Metrics¶

Standard metrics tracked:

likes — post likes
comments — post comments
saves — post saves/bookmarks
shares — post shares
views — post views
impressions — feed impressions
clickThroughRate — CTR
watchTime — video watch time
followerGain — new followers from post

Agent Context Injection¶

The perception summary is automatically injected into the agent's context by ctx-agent when building the memory prelude. This happens when:

CTXDB_PERCEPTION=true (default)
Perception data exists in the workspace

The agent sees the perception summary as part of its context, alongside persona, user profile, and workspace memo content.

Environment Variables¶

Variable	Default	Description
`CTXDB_PERCEPTION`	`true`	Enable/disable perception overlay
`PERCEPTION_MAX_CHARS`	`3000`	Max chars for perception overlay
`PERCEPTION_OUTCOMES_LIMIT`	`20`	Max outcomes loaded for summary
`PERCEPTION_INSIGHTS_LIMIT`	`10`	Max insights loaded for summary
`PERCEPTION_MIN_SAMPLE`	`3`	Min sample size for insight generation

CLI Reference¶

# Record outcome
aios perception record --content-id <id> --platform <name> --content-type <type> [options]

# Generate insights
aios perception insights [--min-sample <n>] [--dry-run]

# View summary
aios perception summary [--format text|json] [--max-chars <n>]

Record Options¶

Option	Required	Description
`--content-id`	Yes	Content identifier
`--platform`	Yes	Platform name (e.g. xiaohongshu)
`--content-type`	Yes	Content type (e.g. note, video)
`--title`	No	Content title
`--publish-time`	No	ISO timestamp
`--snapshot-window`	No	Metrics window (default: immediate)
`--metrics`	No	JSON metrics object
`--context`	No	JSON context object
`--json`	No	Output as JSON

Insights Options¶

Option	Default	Description
`--min-sample`	3	Min outcomes per dimension group
`--dry-run`	false	Preview without storing insights

Summary Options¶

Option	Default	Description
`--format`	text	Output format: text or json
`--max-chars`	10000	Max output characters
`--space`	default	Workspace memory space

FAQ¶

Does Perception automatically publish or optimize content?¶

No. It records outcomes and produces statistical summaries. Publishing, editing, and external actions remain separate workflows with their own approval and verification requirements.

How much history is injected into a task?¶

The summary is bounded by configuration such as PERCEPTION_MAX_CHARS; use the pull-based context rules and request only the material needed for the next decision.

Canonical Docs¶

Pair Perception with ContextDB, Workflow Policy, and Troubleshooting.