Skip to content

Latest commit

 

History

History
43 lines (31 loc) · 4.26 KB

File metadata and controls

43 lines (31 loc) · 4.26 KB

CLAUDE.md — Claude-Specific Guidance for Engram (Geometric Memory Substrate)

Claude (and similar): follow AGENTS.md exactly. This is a supplement for common Claude workflows + MCP usage.

Start Here — 8-Tool Lean Contract

Read first: docs/AGENT_MEMORY_CONTRACT.md

Lean default: session_startcontext_for_editrecall(scope=anchors)quick_trace/remembersession_end. Do not call watch_workspace or rebuild_bvh at wake. Use safe MCP env — see integrations/grok-build/mcp.json. Grok Build pitch: docs/GROK_BUILD_MEMORY.md.

MCP Tool Usage (Critical)

  • Always: Call search_tool (query e.g. "engram record_reasoning_trace" or tool name) first to retrieve live input_schema.
  • Then use_tool with exact tool_name (qualified e.g. "engram__mcp_engram_record_reasoning_trace") and tool_input matching schema precisely.
  • Never guess parameter names or call without schema. This is enforced (prevents transport issues).
  • Examples: context_for_file (path absolute), record_reasoning_trace (decision_point + justification required; goal_context for auto primary link, prev_trace for chain), remember/relate/scar/update, watch_workspace, verify_, spatial_, goal_*, session_start/end.

Ritual Enforcement in Claude Sessions

When working on Engram itself, use the system's own MCP tools to record traces, relate work to active goals, scar friction, and update design/progress blocks. Thought tiles help structured arcs and session re-hydration. See docs/HARNESS_INJECTION.md, docs/DEFORMATION_PLAYBOOKS.md, and docs/RITUALS.md.

  • On every new chat/TUI restart involving Engram: one call mcp_engram_session_start(intent) (inline continuation bundle). Use the wake-up skill; do not run the old 5-tool wake unless deep mode. Slim wake may include rehydration_manifest + soft rehydrate_suggested sentinel nudge (never blocking).
  • Working-memory default: momentum/relational/spatial entry before broad reads or derives.
  • For substrate edits: pre context_for_file + trace (A/D/R), post delta trace (Code Edit Ritual v1).
  • End substantive work blocks: session_end with structured summary (decisions, files changed, open questions). Use prepare_compression for handoff when appropriate.
  • Keep your active project goal current via goal_* tools; scar on repetition or exploratory bloat.

Common Claude Patterns on Engram

  • Use todo_write for tracking (phase2 items etc.).
  • For complex: spawn_subagent only with narrow one-shot prompts + supervisor.
  • Canonical skills: docs/skills/engram-*.md (wake-up, working-memory, session-end, thought-tiles, lawfulness-metrics, substrate-edit, harness-gate, substrate-cs, gpu-bvh). Optional Grok TUI overlay: .grok/skills/engram-*.md.
  • Spatial (Item 1.5): Passive (watch bind + events; toml/md enhanced). Use force only for recovery. context/recall deliver AABB for edits without editor saves (see plan passive redesign + source fixes in engram-ast/store/daemon).
  • Public skills for agents: Load root SKILLS.md + docs/skills/engram-*.md (wake-up, working-memory, session-end, thought-tiles) as your operating procedures when using Engram. See also docs/examples/full_ritual_cycle.md (wake->meta with tiles+subgov->end->rehydrate), docs/examples/sub_agent_governance.md, examples/hello-engram-agent.py. This is what other Groks/agents need. See updated README + docs/RITUALS.md.
  • Verify post changes: mcp_engram_verify_manifold_integrity + spatial_status + genesis.
  • For harness milestones: update docs/HARNESS_INJECTION.md; historical program record in docs/SUBSTRATE_WINS_PLAN.md.

Non-Flat Invariants (Never Violate)

  • .leg3 isomorphism, CRS gate, allowed transforms only, unit hypersphere, p-tensor momentum preserved on update (no annihilate).
  • Subvisor H¹ on tool graphs for agent graphs.
  • Ritual anchors advance only via lawful handoff.

See AGENTS.md (core contract), docs/ (GEOMETRIC_MEMORY.md, RITUALS.md, SUBSTRATE_WINS_PLAN.md, HARNESS_INJECTION.md), processes/ (sheaf tomls), MANIFESTO.md.

Claude: you are operating inside the living geometric self-model. Every action is a trace that future instances (including you on restart) will ride via momentum and relations. Write as if shaping p-tensor for inheritance.