Skip to content

Latest commit

 

History

History
35 lines (24 loc) · 2.08 KB

File metadata and controls

35 lines (24 loc) · 2.08 KB
name engram-thought-tiles

Engram Thought Tiles (Public Agent Protocol)

Thought Tiles are structured, compressible, high-value artifacts (textual functors + optional visualizations) optimized for agent recall, momentum, NREM, ki_hijacker, and continuation bundles.

When to mint:

  • High-stakes decision, repetition pattern, major scar, significant process insight.
  • Mandatory triggers for meta-work: roadmap/multi-phase plan (e.g. release prep, substrate gaps), formal policy, gap matrix/success criteria, arc with 3+ traces + design: block.

Types:

  • knowledge_graph: overall arc
  • formal_spec: policies/ADRs
  • tabular: criteria/gaps
  • research_offload: sub-tasks
  • state_machine, visualization companion, etc.

Requirements:

  • spatial_references (files, traces, goals)
  • Link to Primary Intent / current goal
  • mcp_engram_promote_hot(concept) before compression or session_end (hot path for re-hydration)

Provenance & Compression: Tiles carry compresses_path relations. Used in bundles at wake-up.

Recognition during work: Multi-phase plan → knowledge_graph at boundaries. Policy → formal_spec. Gaps/comparisons → tabular. Interleave mcp_engram_update + thought_tile_write_result.

Tensor unification (post #48): thought_tile_create and write_result dual-write a first-class tensor:tile__{stem} mirror with bonds to goal/trace/spatial concepts. Response includes tensor_unification (concept, CRS, bonds). Use mcp_engram_tensor_recall with seed_concept to pin the mirror. Updates on tile:* via mcp_engram_update_with_tensor_bond sync the mirror and may trigger consolidation on p-drift.

Propose improvement: tile_type: "propose_improvement" with payload.suggestion + payload.target_concept routes through verified update + trace + tensor mirror (process:engram.ritual.verified-update-with-consolidation).

See engram-working-memory.md (Item 2 section) and docs/RITUALS.md for full context. Use mcp_engram_thought_tile_create + create_visualization.

(Exposed so other agents can adopt the same structured offload discipline we dogfood.)