Building RTM Narrative Memory Systems

Encode a narrative into a hierarchical tree and inspect it at multiple abstraction levels:

TypeScript
// 1. Build the tree (K = branching factor, D = max recall depth)
const tree = await tools.call('rtm_create_narrative_tree', {
  text: narrativeText,
  title: "The Lost Kingdom",
  type: "story",
  maxBranchingFactor: 4,   // working-memory chunk size
  maxRecallDepth: 6
});

// 2. Get a high-level (abstract) summary
const abstract = await tools.call('rtm_traverse_narrative', {
  text: narrativeText, title: "The Lost Kingdom", traversalDepth: 1
});

// 3. Get a detailed summary
const detailed = await tools.call('rtm_traverse_narrative', {
  text: narrativeText, title: "The Lost Kingdom", traversalDepth: 4
});

// 4. Model population-level recall variance
const ensemble = await tools.call('rtm_generate_ensemble', {
  text: narrativeText, title: "The Lost Kingdom", ensembleSize: 200
});

Use this checklist when implementing or extending an RTM system (new tool, new corpus, or new deployment):

Progress:
- [ ] 1: Parse narrative into atomic clauses/units (segmentation layer)
- [ ] 2: Choose K (branching factor) and D (max depth) based on text length and desired compression
- [ ] 3: Recursively partition clauses into a tree (rtm-builder) — each node summarizes its children
- [ ] 4: Validate tree statistics (node count, depth, branching distribution) against expected scaling laws
- [ ] 5: Expose traversal at arbitrary depth (rtm-traversal) to retrieve summaries at any abstraction level
- [ ] 6: If modeling human variance, generate an ensemble of trees with randomized partition points (rtm-ensemble)
- [ ] 7: Compute compression ratio = (#leaf clauses) / (#nodes at depth d) for each level
- [ ] 8: If target recall length is given, binary-search traversal depth to match it (findOptimalDepth)
- [ ] 9: Wrap each capability as an MCP tool with strict input validation and typed JSON output

Core algorithm notes

• Tree construction (rtm-builder): Recursively split the clause sequence into ≤K contiguous chunks; each chunk becomes a child subtree. Stop recursion at D levels or when a chunk is a single clause. Each internal node stores a generated/extracted summary of its subtree's text span.
• Traversal (rtm-traversal): To recall at depth d, walk the tree and collect node summaries at level d (or leaves, whichever is shallower) — this is the "clause recall sequence." Concatenating these summaries in order reconstructs a depth-appropriate retelling.
• Ensemble (rtm-ensemble): Fix the text but randomize partition boundaries (within K/D constraints) across N tree instances. Aggregate statistics (mean/variance of node count, compression ratio, recall-sequence length) approximate population-level recall differences.
• Optimal depth search (findOptimalDepth): Since recall length grows roughly geometrically with depth (~K^d leaves per level), use the relation depth ≈ log_K(targetRecallLength / rootBranches) as a starting estimate, then refine by direct traversal-length comparison at adjacent depths.
• Scale invariance: For long narratives, compression ratio between adjacent depths should stabilize (approach a constant), reflecting the K-ary branching; deviations signal uneven clause density and may require re-segmentation.

Example 1: Creating a tree Input:

JSON
{"text": "Once upon a time... [500-word story]", "title": "The Lost Kingdom", "type": "story", "maxBranchingFactor": 4, "maxRecallDepth": 6}

Output:

JSON
{
  "treeId": "rtm_a1b2c3",
  "stats": {"totalNodes": 85, "leafCount": 64, "depth": 6, "avgBranching": 3.8},
  "compressionRatio": {"depth1": 0.015, "depth3": 0.12, "depth6": 1.0}
}

Example 2: Depth-controlled traversal Input: traversalDepth: 1 on the above tree Output:

JSON
{
  "summaries": ["A wise king protects his kingdom from a growing threat."],
  "recallSequence": ["A wise king protects his kingdom from a growing threat."],
  "compressionRatio": 0.015
}

Example 3: Optimal depth search Input: targetRecallLength: 50 clauses Output:

JSON
{"optimalDepth": 4, "actualRecallLength": 48, "accuracy": 0.96}