name	engram-working-memory
description	Runtime discipline during work: context_for_edit before edits, anchor recall before derive, quick_trace at forks, update over forget+remember.
metadata	{"short-description":"Work loop — edit context, recall, trace"}

Engram Working Memory

Trigger: After wake, for every edit and decision during the session.

Before editing a file (preferred safe composite)

mcp_engram_safe_edit_and_verify(
  path="/absolute/path/to/file",
  decision="What you plan to change",
  why="Justification",
  arc_delta="delta: narrative after edit (optional)",
  goal_context="goal:..."
)

Few-shot (1): {"path":"/home/user/Engram/crates/engram-server/src/mcp.rs","decision":"Add safe_edit composite tool","why":"Agent tool fidelity goal — one-shot verified edit path","arc_delta":"delta: registered mcp_engram_safe_edit_and_verify handler","goal_context":"goal:agent_tool_fidelity_v1"}

Few-shot (2): {"path":"/home/user/Engram/docs/AGENT_MEMORY_CONTRACT.md","decision":"Refresh 8-tool examples","why":"Mirror hardened few-shots in docs","run_verify":true}

Or lean pre-edit only: mcp_engram_context_for_edit(path="...") — /engram-edit or /engram-safe-edit

Before heavy reasoning

mcp_engram_recall(query="<goal or trace keywords>", scope="anchors", k=5)

Or: /engram-recall

At forks

mcp_engram_quick_trace(decision="...", why="...", goal_context="goal:...")

Or: /engram-trace

Writes (Layer 1 — see TOOL_DECISION_MAP)

mcp_engram_recall — always first
Score >0.85 → mcp_engram_update_with_tensor_bond (preferred) or mcp_engram_update (only legal mutation of existing concepts)

Few-shot update_with_tensor_bond (1): {"concept":"mcp__fn__dispatch__arc","new_text":"delta: wired safe_edit handler","recall_query":"mcp dispatch edit arc","bond_label":"edit_fidelity"}

Few-shot update_with_tensor_bond (2): {"concept":"design:agent_tool_fidelity_v1","new_text":"Phase 1: composite tools shipped","recall_query":"agent tool fidelity","scar_on_mismatch":true}

Few-shot remember (1): {"concept":"harness:agent_tool_fidelity_v1","text":"Deterministic suite for edit/update tool fidelity >=95%."}

Few-shot remember (2): {"concept":"user__prefers_absolute_paths","text":"Always pass absolute paths to context_for_edit and safe_edit_and_verify."}

No match → mcp_engram_remember
Verified fix → mcp_engram_remember_solution
Dead end / doom loop → mcp_engram_scar
Chain → mcp_engram_relate to goal/trace

Never forget + remember on the same concept.

Post-edit reflection (mandatory after substantive change)

Execute reflection_suggested from composite responses: quick_trace delta → verify_block_lawfulness (merkle) → tensor:edit_pattern_* upsert.

Rituals: process:engram.ritual.safe-code-edit, process:engram.ritual.verified-memory-update

Harness gate: tools/test-harness/bin/engram-harness.sh --suite agent-tool-fidelity

name	engram-working-memory
description	Runtime discipline during work: context_for_edit before edits, anchor recall before derive, quick_trace at forks, update over forget+remember.
metadata	{"short-description":"Work loop — edit context, recall, trace"}