session-memory（会话内存）

Session Memory (MANDATORY)

Overview

CC10X memory is a small, stable, versioned Markdown index plus durable workflow artifacts. It exists to survive compaction, preserve decisions, and keep BUILD / DEBUG / REVIEW / PLAN aligned across sessions.

Memory is an index, not a transcript. Be brief. Distill decisions, learnings, references, and verification evidence into durable, reusable notes.

Reference Map

Keep this file as the control plane. Read only the reference you need:

• references/memory-model-and-ownership.md for memory surfaces, ownership, promotion, and workflow markers
• references/memory-operations.md for permission-free operations, router-only persistence, and targeted edit patterns
• references/memory-file-contracts.md for required headings, stable anchors, templates, and auto-heal rules
• references/context-budget-and-checkpointing.md for context-budget rules, warning signs, and checkpoint triggers

The Iron Law

EVERY WORKFLOW MUST:
1. LOAD memory at START and before key decisions
2. EMIT memory-worthy notes before finishing
3. LET the router/workflow finalizer persist markdown memory

If memory is skipped, the workflow becomes conversation-dependent. If memory is narrated instead of distilled, it bloats context and loses signal.

Load-Bearing Boundaries

• Memory lives under .claude/cc10x/v10/.
• The router loads and auto-heals memory files before routing or resume.
• WRITE agents read memory, but do not edit .claude/cc10x/v10/*.md directly.
• WRITE agents emit structured MEMORY_NOTES in their Router Contract.
• READ-ONLY agents emit ### Memory Notes (For Workflow-Final Persistence).
• The router-owned memory-finalize task is the only final writer of memory markdown files.
• Workflow artifacts under .claude/cc10x/v10/workflows/{wf}.json remain the durable execution truth; markdown memory is the stable index.
• Required headings and anchors are a contract. Do not invent new file layouts or marker schemes.
• Router-owned markers such as [DEBUG-RESET: wf:{...}] and [cc10x-internal] memory_task_id are read and respected by agents; agents do not invent replacements.

Memory Surfaces

Use the right layer for the right kind of information:

• activeContext.md: current focus, recent changes, decisions, learnings, references, blockers
• patterns.md: reusable project standards, gotchas, conventions, and skill hints
• progress.md: current workflow, tasks snapshot, completed items, verification evidence
• docs/plans/* and docs/research/*: the detailed artifacts; memory points to them
• .claude/cc10x/v10/workflows/{wf}.json and .events.jsonl: the durable orchestration truth

Read references/memory-model-and-ownership.md if you need the full ownership model or the promotion ladder.

Distillation Rule

Memory entries must be distilled, not narrated.

Preserve:

• decisions with rationale
• causal learnings
• stable module boundaries and file paths
• verification commands and exit truth
• external artifact paths the next session will need

Strip:

• decorative prose
• step-by-step diary text
• speculation that was never resolved
• redundant retellings of the prompt
• unstable line numbers unless they are the only durable locator available

Test:

If a fresh agent reading memory cannot reconstruct the current constraints and next move without re-reading the whole conversation, the memory is under-distilled.

What To Load

Always Load

At workflow start, continuation, or resume, read all three memory files:

.claude/cc10x/v10/activeContext.md
.claude/cc10x/v10/patterns.md
.claude/cc10x/v10/progress.md

Re-Read Before These Actions

File Selection Matrix

What To Persist

Emit only memory-worthy items:

• decisions that change implementation direction
• learnings that prevent repeating the same mistake
• verification evidence with commands and exit truth
• deferred non-blocking issues that should survive the session
• plan, design, or research references the next workflow must follow
• clarified user standards or conventions that should become durable project patterns

Do not persist:

• whole diffs
• verbose execution logs
• celebratory narration
• "looked correct" claims without evidence
• duplicate notes that add no new constraint

Agent Output Contract

WRITE agents persist memory through the Router Contract:

MEMORY_NOTES:
  learnings: ["Key causal insight"]
  patterns: ["Reusable gotcha or convention"]
  verification: ["`npm test` -> exit 0"]
  deferred: ["Non-blocking follow-up or risk"]

Rules:

• Keep each entry one line when possible.
• Prefer stable nouns over temporary wording.
• Put verification truth in verification, not in narrative prose.
• Put non-blocking follow-up work in deferred, not in ad-hoc tasks.
• If a complementary skill failed to load or a durable artifact was missing in a way that matters, note it in learnings or deferred.

READ-ONLY agents use their own ### Memory Notes (For Workflow-Final Persistence) section, but the router persists both shapes into the same final memory update.

Project Skill Hints

After loading memory, check patterns.md ## Project SKILL_HINTS.

• Invoke listed skills after memory load when they are relevant and installed.
• Treat missing skills as non-fatal; note the miss in memory-worthy output if it changed the workflow.
• Do not invent or rewrite skill IDs from this skill.
• Do not try to "maintain" ## Project SKILL_HINTS from a write agent by editing memory directly. If a new durable recommendation matters, surface it in planner output or memory notes for router-owned persistence.

Workflow Pattern

Start

1. Read all three memory files.
2. Read current plan, design, or research files from activeContext.md ## References only if the workflow needs them.
3. Load approved project/domain skills from ## Project SKILL_HINTS after memory is loaded.
4. If context is already getting heavy, apply references/context-budget-and-checkpointing.md.

During

• Re-read memory before key decisions, not after the mistake.
• Respect router-owned markers such as [DEBUG-RESET: wf:{...}].
• Keep notes distilled enough to survive compaction.
• Prefer durable artifacts and file paths over long conversational summaries.

End

1. Distill decisions, learnings, verification, and deferred items into MEMORY_NOTES.
2. Keep each note factual and reusable.
3. Assume the router/workflow finalizer will persist the markdown files.
4. Do not report the workflow as truly complete if the memory cycle is missing.

Context Budget And Checkpointing

Long sessions degrade before they fail loudly.

Use references/context-budget-and-checkpointing.md when:

• debugging has gone through multiple failed attempts
• planning/research is pulling in many large files
• the session has already accumulated many tool calls
• you notice vagueness, skipped protocol steps, or rising uncertainty

Checkpoint principle:

• persist durable state into workflow artifacts, plans, research files, and memory notes
• do not rely on the conversation transcript to carry load-bearing context

Red Flags

Stop and correct course if you catch yourself:

• starting work without loading memory
• making decisions without checking ## Decisions or project patterns
• writing long diary-style memory notes
• claiming you will "remember later"
• editing .claude/cc10x/v10/*.md directly from a write agent
• inventing new headings, markers, or memory file shapes
• treating stale conversation context as better than durable memory

Rationalization Prevention

Verification Checklist

• Memory loaded at workflow start
• Relevant decisions/patterns checked before new decisions
• Notes distilled into MEMORY_NOTES
• Verification truth captured in memory-worthy form
• No direct write-agent edits to .claude/cc10x/v10/*.md
• Context-budget checkpointing considered if the session got heavy

Cannot check these boxes? The memory cycle is incomplete.