Memory Palace Curation Workflow

Learn how the research interceptor collaborates with knowledge-intake for effective curation.

Prerequisites

• Memory Palace plugin installed
• Understanding of cache modes (see Cache Modes)

Objectives

By the end of this tutorial, you’ll:

• Understand the intake flag payload
• Process the intake queue
• Use dual-output workflows
• Maintain curation quality

The Intake Flow

When a research query runs, Memory Palace evaluates whether new information should be captured:

Query Execution
      |
      v
[Hook Evaluation]
      |
      +-- Build IntakeFlagPayload
      |   - should_flag_for_intake
      |   - novelty_score
      |   - domain_alignment
      |   - duplicate_entry_ids
      |
      v
[Decision]
      |
      +-- High novelty + domain match --> Flag for intake
      |
      +-- Low novelty or duplicate --> Skip intake
      |
      v
[Output]
      - Telemetry row
      - Queue entry (if flagged)

Step 1: Understanding IntakeFlagPayload

The IntakeFlagPayload dataclass tracks three signals:

# memory_palace/curation/models.py
@dataclass
class IntakeFlagPayload:
    should_flag_for_intake: bool
    novelty_score: float
    domain_alignment: List[str]
    duplicate_entry_ids: List[str]
    intake_delta_reasoning: str

Step 2: Monitoring the Hook

The hook outputs intake context to the runtime transcript:

[Memory Palace Intake]
Novelty: 0.75
Domains: python, async
Duplicates: []
Flag: True
Reasoning: High novelty content in aligned domain

Step 3: Processing the Intake Queue

When should_flag_for_intake=True, the hook writes to:

data/intake_queue.jsonl

Process the queue:

# View pending items
cat data/intake_queue.jsonl | jq .

# Process with CLI
uv run python skills/knowledge-intake/scripts/intake_cli.py \
  --queue data/intake_queue.jsonl \
  --review

Step 4: Intake Decision Options

For each queued item:

# Accept item
uv run python intake_cli.py --item abc123 --accept

# Merge with existing
uv run python intake_cli.py --item abc123 --merge entry-456

# Reject
uv run python intake_cli.py --item abc123 --reject

Step 5: Using Dual-Output Mode

Generate both palace entry and developer documentation:

uv run python intake_cli.py \
  --candidate /tmp/candidate.json \
  --dual-output \
  --prompt-pack marginal-value-dual \
  --auto-accept

This creates:

1. Palace entry: Stored in corpus
2. Developer doc: Added to docs/
3. Prompt artifact: Saved to docs/prompts/<pack>.md

Step 6: Telemetry Review

Check telemetry for intake patterns:

# View recent decisions
tail -20 data/telemetry/memory-palace.csv

Columns include:

• novelty_score
• aligned_domains
• intake_delta_reasoning
• duplicate_entry_ids

Curation Best Practices

Regular Review Cadence

1. Daily: Check intake queue
2. Weekly: Review telemetry patterns
3. Monthly: KonMari session (prune low-value entries)

Document Decisions

Update docs/curation-log.md after each session:

## 2025-01-15 Curation Session

### Promoted
- async-patterns.md: High usage, evergreen

### Merged
- context-managers.md + async-context.md: Redundant

### Archived
- python-3.8-features.md: Outdated

Maintain Vitality Scores

Keep data/indexes/vitality-scores.yaml current:

entries:
  async-patterns:
    vitality: evergreen
    last_accessed: 2025-01-15
  python-3.8-features:
    vitality: probationary
    last_accessed: 2024-06-01

Troubleshooting

Too many intake flags

Solution: Raise intake threshold

# In config
intake_threshold: 0.7  # Higher = fewer flags

Missing domain alignment

Solution: Update domains of interest

# hooks/shared/config.py
domains_of_interest:
  - python
  - async
  - testing
  - architecture

Duplicate detection failing

Solution: Rebuild indexes

uv run python scripts/build_indexes.py

Verification

Confirm the workflow:

1. Make a research query
2. Check telemetry for decision
3. View intake queue if flagged
4. Process queue item
5. Verify corpus update

# After processing
ls docs/knowledge-corpus/ | grep new-entry

Next Steps

• Return to Plugin Overview
• Explore Capabilities Reference