Skip to content

GROK_BUILD_MEMORY.md

Latest commit

 

History

History
113 lines (81 loc) · 5.38 KB

GROK_BUILD_MEMORY.md

File metadata and controls

113 lines (81 loc) · 5.38 KB
Raw

Engram for Grok Build — The Memory Layer Agents Actually Use

One-line pitch: Persistent geometric memory over MCP — one-call wake, anchor-first recall, edit-scoped spatial context, structured handoff. Runs local. Survives 200k-block stores without OOM.

Why xAI should care

Flat RAG (vectors + chunks) gives agents retrieval. Engram gives agents continuity:

Flat memory Engram
Similarity search over chunks Goals, traces, scars, rituals as first-class anchors
Session dies → context lost session_end → structured handoff → next session_start rehydrates
"Remember this" = embed + store CRS-gated blocks + Merkle lineage + update (no annihilate)
Code = grep/RAG Code atlas v2 — context_for_edit: AST + __arc + traces/scars at locus
No trust model CRS tiers, scars, lawfulness verify

The Grok Build integration story: Engram is already an MCP server. Grok Build spawns it once per workspace. Agents follow an 8-tool contract — not all 79 tools every session, not a 5-tool wake cathedral.

Native tools (e.g. scheduler_create for Grok's /loop recurring prompts) must be called bare/direct (never through use_tool or Engram MCP wrappers). Use /engram-loop or the equivalent ritual for Engram-aware /loop handling: parse per the spec, bare native call, immediate Engram quick_trace + remember/relate (job id to goal/tile/process), subvisor governance, honest confirmation or scar on native format error. See grok-plugin-engram/commands/engram-loop.md. This prevents the historical "doom loop" of misrouted native calls + false success claims.

The 8-tool contract (ship this in Grok Build docs)

WAKE   → session_start(intent)              # inline continuation + harness_injection.suggested_actions
WORK   → context_for_edit(path)             # before editing a file
       → recall(query, scope="anchors")     # goals/traces when stuck
       → quick_trace / remember             # forks and facts
END    → session_end(summary)               # handoff packet for next session
PROBE  → get_backend_readiness              # lean vs deep, RSS-safe mode
MODE   → set_memory_mode("deep")            # only when full recall needed

Load for every agent: docs/AGENT_MEMORY_CONTRACT.md (8-tool highway) + docs/TOOL_DECISION_MAP.md (full 79-tool map) + docs/CODE_ATLAS_CONTINUITY.md (situated edit memory) + docs/DEFORMATION_PLAYBOOKS.md (JIT RSI) + SKILLS.md

Recommended MCP config (all ecosystems)

Use scripts/engram-grok — sets ENGRAM_PROFILE=agent (lean CUDA, deferred BVH, anchor recall):

{
  "mcpServers": {
    "engram": {
      "command": "/path/to/Engram/scripts/engram-grok",
      "args": ["mcp"],
      "env": {
        "ENGRAM_STORE": "~/.engram/stalks/",
        "ENGRAM_PROFILE": "agent"
      }
    }
  }
}

See integrations/README.md for Cursor, Claude, Antigravity, Codex.

Validated on: 183k+ .leg blocks, ~230MB RSS, <600ms lean wake (harness), transport stable. agent-memory suite green including harness_injection queue.

What NOT to teach agents (lean mode)

These tools still exist for power users — do not put them in the default Grok Build prompt:

Tool Why avoid in default
watch_workspace Was 40GB RAM on large repos; deferred by default
rebuild_bvh Minutes + GB RAM; opt-in quality mode only
get_continuation_bundle Redundant — inline in session_start
list / list_concepts Full store scan on 100k+ blocks
query_with_momentum at wake Extra round-trip; use anchor recall first
summarize at wake Duplicates inline bundle

See MCP_TOOLS_REFERENCE.md for full tier list.

Differentiators vs mem0 / Letta / vector DBs (for README hero)

  1. Hardware-native blocks — 256KB .leg3, O_DIRECT NVMe, optional GPUDirect
  2. Non-flat geometry — q/p tensors, CRS Lyapunov, momentum recall
  3. Rituals as hygiene — scar, verify, trace chains, session handoff
  4. Situated code atlas — structure + update(__arc) edit continuity + traces at file:line (CODE_ATLAS_CONTINUITY.md)
  5. Declarative process sheaf — processes/*.toml, subvisor H¹ governance
  6. MCP-native — 8-tool lean path + power tools for depth

Shipped in Agent Memory MVP

  • Core loop (A1–A6): one-call wake, lean/deep mode, anchor recall, context_for_edit, structured handoff
  • docs/AGENT_MEMORY_CONTRACT.md — canonical 8-tool agent entry
  • SKILLS.md — points to contract first
  • Grok Build plugin — 20+ /engram-* slash commands + MCP spawn per workspace
  • docs/TOOL_DECISION_MAP.md — full 79-tool decision map
  • docs/DEFORMATION_PLAYBOOKS.md — JIT RSI playbooks
  • integrations/ — MCP config templates for Cursor, Claude, Codex, Antigravity

Roadmap (post-MVP)

  • Async load_process_sheaf (faster session_start)
  • note() write primitive (remember/update unified)
  • Harness gate in CI
  • Optional: mark MCP tool descriptions with [ESSENTIAL] / [POWER] / [LEAN:AVOID] in mcp.rs

Do not delete MCP tools — tier them in docs; removal breaks power users and TUI paths.