memory-continuity

name	memory-continuity
description	Preserves cross-session continuity with deterministic MEMORY.md, index.md, and snapshot handoffs under .agentic/memories/.
version	1.0.0
tags	["agentic","memory","continuity","workflow"]
resources	["memory-current-template.md","index-template.md","memory-template.md"]
vendor_support	{"claude":"native","opencode":"native","copilot":"prompt-inject","codex":"prompt-inject","gemini":"prompt-inject"}

Memory Continuity Skill

Use this skill to preserve high-value context between clean sessions while avoiding noisy or low-signal carryover.

Storage Layout (Hybrid Contract)

Use this deterministic structure:

.agentic/
  memories/
    MEMORY.md
    index.md
    snapshots/
      YYYY-MM-DDTHH-MM-SSZ.md

Rules:

• If .agentic/memories/ does not exist, create it and required child paths.
• MEMORY.md is the current canonical continuity file.
• index.md is a small pointer log to snapshots.
• snapshots/ stores timestamped continuity checkpoints.
• Use memory-current-template.md for MEMORY.md initialization.
• Use index-template.md for index.md initialization.

First-Run Initialization and Bootstrap Rules

When MEMORY.md and index.md do not exist:

1. If there is active iteration context (for example an open issue/PR, active branch work, or already-decided next steps), initialize both files from templates immediately.
2. If there is no active iteration context, inspect commit history for signal:

• if history is rich enough to derive meaningful continuity, ask the user before generating a brief bootstrap memory from commits;
• if history is not rich enough, ask the user before leaving memory uninitialized.

3. Never invent synthetic history. If signal is insufficient, state that explicitly.

Start-of-Session Retrieval Workflow

1. If files are missing, apply First-Run Initialization and Bootstrap Rules.
2. Read .agentic/memories/MEMORY.md if present.
3. If index.md exists, read only the latest 1 to 3 snapshot pointers.
4. Extract only high-signal items:

• current objective
• decisions made and rationale
• open risks/questions
• next concrete steps

5. Do not import verbose logs, full transcripts, or resolved low-impact details.

End-of-Session Update Workflow

1. Create directories/files if missing:

• .agentic/memories/
• .agentic/memories/snapshots/
• .agentic/memories/MEMORY.md (from memory-current-template.md)
• .agentic/memories/index.md (from index-template.md)

2. Build a new memory entry using memory-template.md.
3. Update MEMORY.md additively:

• keep existing unresolved items
• append new high-signal updates
• remove or compress resolved/noisy items

4. Write a timestamped snapshot:

• filename format: YYYY-MM-DDTHH-MM-SSZ.md

5. Append one row in index.md pointing to the snapshot and objective.

Multi-Agent and Worktree Coordination Rules

• Treat MEMORY.md as shared state: do not delete or rewrite another agent's active workstream row or unresolved items.
• Each agent should update only its own workstream row (keyed by worktree + branch + agent label) and append decisions/risks with timestamps.
• If multiple agents are active, preserve all active workstream rows and resolve conflicts by appending a new decision note rather than overwriting history.

Signal-Over-Noise Constraints

Always exclude:

• full command history
• raw terminal dumps and repetitive logs
• obvious or already-resolved implementation minutiae
• speculative ideas with no immediate decision/value

Prefer compact, testable statements and explicit decisions.

Pruning and Compaction Rules

• Keep MEMORY.md under ~120 lines by compacting resolved sections.
• Keep only the latest 20 snapshot files by default.
• Preserve older snapshots only if tagged as milestone decisions.
• On compaction, never drop unresolved risks or agreed next steps.

Output Requirements

When this skill runs, report:

• files created or updated
• snapshot filename written
• whether first-run bootstrap was used (and whether user approval was requested)
• whether compaction/pruning occurred
• any unresolved risks carried forward