feat(prep): GitHub MVP - non-flat geometric memory, rituals, passive spatial AABB, agent skills exposure, subvisor governance#27
Merged
Conversation
- Add `embed_w: Option<Vec<f32>>` + `embed_src_dim: usize` to StoreHandle - `load_embed_w()` reads W.bin (2816×8192, 92.3MB) from ENGRAM_EMBED_W_PATH or ~/Documents/CodeLand/data/models/embed_projection_W.bin (default) - `try_project_text()`: GET /v1/embeddings → matmul W → U(1)^8192 complex phase - `remember()`: replaces Helical Baptism q-vector with projected vector if server is reachable and W is loaded; marks block.payload with [CAL] tag for observability - Graceful fallback to Helical Baptism if embedding server is down or W not loaded - Euler characteristic gate still validates all vectors (projected or Baptism) Closes Phase 111-A (W.bin RMSE=0.026647) + Phase 111-B Encoding Commutativity Gap between Engram and Oracle strata is now closed.
… tiles for leg-browser (codeland-integration goal)
- Complete source: engram-server MCP (50 tools), engram-core, engram-gpu with Metal backend for Apple Silicon - Ritual skills in .grok/skills/ (engram-wake-up, working-memory, session-end, goal) - scripts/leg launcher, tools/leg-browser, all integrations and docs - .gitignore already protects private *.leg3 / manifold data - Ready for build on aarch64-apple-darwin with existing Metal + CPU fallbacks This commit snapshot turns the EngramGrok private repo into the actual cloneable workspace instead of docs-only. Created under ritual discipline after explicit user request to deliver the entire thing, not just .md files.
…gent audit) - DEPLOYMENT_MODES.md: Replace all '15 tools' language with current 50+ surface reality + mandatory search_tool first / use_tool rule + Rule 6 + trace/scar binding. - README.md: Update '31 Tools' header and add explicit MCP hygiene, Rule 6, and scar/trace mandates at the top of the reference section. Qualify query_with_momentum description. High-priority fixes from the first completed subagent report (Tool usage/hygiene/MCP patterns). Other 4 alignment auditors still running. These changes will be pushed to the private EngramGrok handoff workspace so the Mac Mini 4 clone reflects actual 2026 practice. Under full engram-working-memory + Code Edit Ritual discipline.
…ac Mini use case - New root HOW_WE_ACTUALLY_USE_THIS_IN_2026.md (warm, honest current 2026 ritual loop, honest state of leg-browser/dual tiles/goals as emerging, loud private data separation, explicit Mac/Apple Silicon notes, points to living .grok/skills/ as runtime truth). - Root README: prominent 'New here?' box at the very top pointing to the new doc; cleaned the leg-browser callout (removed specific goal/ritual IDs from this wave, set realistic expectations for STATIC vs dynamic, kept the one-command launcher as the practical daily driver). These are the first high-leverage deliverables from the full five-agent documentation alignment audit. Internal jargon, outdated assumptions, and pre-work artifacts are being systematically addressed so the private EngramGrok clone (the full workspace) is actually usable for a non-technical spouse on a maxed-out Mac Mini 4 with Grok Build. All changes pushed to the handoff remote.
…, workflow, honest state, Mac notes, private separation, rituals — removed explicit wife/personal framing per request)
…itual legominism rehydration + precise launch tracking - docs/Engram_Build_Launch_Configuration.md: new single living source of truth for exact build/launch commands, observed runtime values (ENGRAM_KI_ARTIFACTS_DIR, ki_hijacker path, CUDA/Sheaf, ego.leg3), rollback history, and correction rules. Updated on every state transition per user demand after TUI restarts. - .grok/skills/engram-wake-up/SKILL.md + engram-session-end/SKILL.md: symmetric legominism rehydration logic (search_by_relation + compresses_path + high-fruit + serves to codeland goal 1780091465... + NREM/ego.leg3 synthesis tiles). High-lineage Phase 3 closure artifacts (Track 1/2 formal_spec, external-agent guide, handoff, MVP showcase) now auto-surface on TUI restart. - crates/engram-gpu/src/bvh.rs + backend.rs: temporary large-manifold guard (n>100k returns valid empty BvhManifold + WARNING + DIAG; 128 MiB dedicated stack thread for recursive build_top_down). Unblocks 154k+ concept CUDA path while post-LBVH segfault root cause is isolated. - README.md: dual-track charter surfaced (deepest TUI ritual self-model demonstration + neutral memory utility/BYOP for external agents, "use is more important than understanding", Phase 3 assets as lead public story). User confirmed: stable binary from last GitHub push (ac3509a) + TUI restart succeeded. This is now the daily driver (~/.engram-ac3509a9/bin/engram). Dev branch remains guarded for diagnosis. Ties directly to living coordination tile (knowledge_graph_phase-1-cross-workstream-coordination--ws1-hot-p), formal_spec_track-1-plan, formal_spec_track-2-plan, PGFS v0.1, orchestration_autonomy_guidelines_v0_1, and Phase 3 handoff. MCP transport closed during this agent's wake-up ritual in CLI context (substrate signal recorded in config); user's TUI has the live connection and active rehydration. Refs: Track 1/2 plans, codeland goal, 2026-06 roadmap, autonomy guidelines 9-item checklist.
…giene - Passive spatial: engram-ast now supports .toml (process tables/sections) + enhanced .md (frontmatter, headings, code fences) with AABB. Daemon/store: auto initial on watch_bind + fs events, robust ignores (CWD .engramignore + defaults), context/recall fallback to regular fetch (fresh items visible), state provlog hygiene. No more "user open+save" or bootstrap_user_action. - Public MVP: README hero/table/comparison + Memory Model, Cargo metadata (geometric/sheaf/rituals/256KB), .github/ (matrix CI + harness/ritual job, PR/ISSUE templates with hygiene checklist), AGENTS.md/CLAUDE.md (ritual contract), CHANGELOG, new docs/GEOMETRIC_MEMORY.md RITUALS.md MCP_TOOLS_REFERENCE.md + updated GITHUB_MVP_PREP_PLAN.md (full execution + redesign), examples/ (sanitized runnable), processes/ (first-class declarative toml sheaf). - Hygiene/sanitization: .gitignore tightened (.grok/, data/, internal conv/handoff); all public surface double-checked for paths (placeholders only in prep notes); current build (target/debug/engram) throughout; path loader portable. - Dogfood: engram-wake-up/working-memory, pre/post traces, spatial calls, mcp_update, relate to goal:1780419540, verify healthy, parser unit tests (8 toml + 20 md items on real files). - Closes: the "spatial-user-action nonsense" item; enables reliable AABB for continuation, Code Edit Ritual, self-edits. See docs/GITHUB_MVP_PREP_PLAN.md for phases, recon, success, redesign details.
…gents/Groks - Root SKILLS.md: top-level index for easy discovery, linking public skills/, full-cycle demo, sub-agent gov, hello script, processes/, etc. - examples/hello-engram-agent.py: tiny self-contained runnable demo that loads docs/skills/*.md and walks one complete wake->meta-work(with tiles)->session-end->rehydrate loop using client shim. - docs/examples/full_ritual_cycle.md: rich cookbook-style full-cycle demo (wake -> heavy meta with tiles + explicit sub-agent governance call -> end with COMPRESS/handoff -> next wake rehydrate/continue), with runnable snippets. - docs/examples/sub_agent_governance.md: dedicated patterns for sub-agents (subvisor H¹ on tool graphs, narrow one-shot prompts, supervisor monitoring/kill/fallback, escalation via helpers:meta_work_escalation_v1 + current_meta_arc, scars/traces from real prep doom-loop lessons, best practices). - Updated public surface: README (prominent agent section + SKILLS.md), docs/RITUALS.md (cross-refs), docs/GITHUB_MVP_PREP_PLAN.md (execution log for this pass), AGENTS.md/CLAUDE.md (load instructions). - All dogfooded (trace recorded + relate to primary goal), paths sanitized (only meta notes in plan), build clean, no private leaks in public commit set. - Addresses the "missing items" for agents to have the exact skills/rituals we use, plus richer demos and gov patterns. See SKILLS.md and docs/examples/ for agent onboarding. Part of GitHub MVP prep public face.
…HUB_MVP_PREP_PLAN.md Small follow-up to keep the living prep plan up to date with the public surface additions for agents. No other changes.
…doff push and PR #6 log Follow-up to the skills exposure commits. Keeps the living plan current for the push to EngramGrok repo as specified. All public, sanitized.
…e Leg3Pointer import, SymplecticState move, incomplete Memory init) - Removed re-import of Leg3Pointer (already in unconditional use). - Split SymplecticState import under #[cfg(target_os = "macos")] to avoid unused on other builds. - Extracted lens before move in promote_geo_snapshot_to_high_priority. - Populated all Memory fields (AABB, alpha, l2, energetics, zedos, etc.) from HolographicBlock in query scoring path (matches bvh.rs style for consistency). Fixes the exact E0252/E0382/E0063 from aarch64-apple-darwin CI after engram-gpu metal cfg build. Dogfooded: spatial context/recall attempted + trace:1780429158... + relate to goal:1780419540 + verify healthy. Part of pushing the full prep (skills exposure + spatial passive) to EngramGrok public handoff.
- Fixed engram-gpu metal_backend compile errors (imports, move, Memory fields). - Commit dacabd3 for the fix. - Re-pushed the branch to EngramGrok public handoff. - Updated plan with details. Dogfooded via engram trace + relate to goal. Part of making the full MVP (skills + spatial + now build hygiene) public on EngramGrok.
…t, doc_lazy_continuation, needless_range_loop, io_other_error, let_unit_value, explicit_auto_deref) + borrow-safe residual loop - Add Default impl for SymplecticState (delegates to new()). - Indent continuation doc list items in ops.rs. - Replace needless range loops with iterators/zip/enumerate (ops normalize_in_place, backend residual). - Use Error::other and omit unit let-binding in async storage. - Remove explicit * deref in backend query paths. All per Clippy suggestions. Verified clean cargo check -p engram-core. Part of stabilizing the EngramGrok public handoff after CI failures.
Ship ENGRAM_PROFILE=agent|deep|ui|dev to replace env-var soup for local agent MCP. Consolidate launcher (scripts/engram-grok), minimal grok-build MCP config, and backend_readiness/handoff profile fields. Agent memory MVP (T0–T2): - profile.rs with apply() defaults for lean CUDA, deferred BVH, anchor recall - session_start inline continuation bundle + warm_wake_anchors at wake - context_for_edit enriched JSON (related_goals/traces, profile, memory_mode) - MCP param aliases (goal_id, from/to/relation) and clearer errors - agent-memory harness suite with profile + context_for_edit assertions Tier 3 recall on 181k stores: - anchors scope no longer backfills broad overview pool (was ~2k blocks) - ENGRAM_LEAN_ANCHOR_POOL default 800; high-priority fetch in scoring - recall(anchors) p50 ~0.9s (down from ~2.3s) on sampled_bounded path Docs: AGENT_MEMORY_CONTRACT, GROK_BUILD_MEMORY, MVP_GAP_CLOSURE_PLAN, consolidation plan, updated skills and GitHub MVP prep surfaces. Note: restart Grok TUI after pull to pick up engram-grok launcher.
Git hygiene: - gitignore harness logs/results, __pycache__, *.pyc - remove tracked conv_task/conv_arc internal docs (already gitignored pattern) - remove 70 harness run artifacts from index Unified integrations (Grok, Cursor, Claude, Antigravity, Codex, local): - integrations/README.md master guide + mcp.engram.template.json - integrations/codex/README.md for CLI agents - All MCP configs: scripts/engram-grok + ENGRAM_PROFILE=agent (2 env vars) - Portable engram-grok (repo-relative binary resolution, no hardcoded paths) - system-prompts/README.md: 8-tool lean contract (replaces watch+summarize wake) - README, FIRST_RUN, GROK_BUILD_MEMORY, DEPLOYMENT_MODES, AGENTS aligned
…lence cfg warnings
…; wgpu HotBlockCache borrow (use Arc clone for &Device)
…x E0382 use-after-move for scores_buf
…pport (P1–6 + mixed arc)
…ed/cfg/parens via main.rs allows + async-io feature in Cargo.toml + safe fixes); clippy --workspace run (many lints addressed: ast, core range/backfill, gpu cstr/dead/doc, build flag/if_same, etc); pre-push per CONTRIBUTING. Includes toml feature, main allows (unused_mut, unnecessary_parens, type_complexity, ptr_arg), core/ast/gpu/backfill fixes. On top of 6025ced + v0.4.1.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR brings the current Engram MVP to public GitHub after extensive prep, dogfooding with rituals, and addressing agent usability.
Key changes (from GITHUB_MVP_PREP_PLAN.md execution)
Ritual hygiene followed throughout
Post-push
See full plan + execution in docs/GITHUB_MVP_PREP_PLAN.md. All under rituals, MCP, current build.
Ready for review/merge. (Created after explicit 'push it public'.)