Memory System

Claude Code’s memory system gives the agent persistent, file-based knowledge that survives across sessions. Unlike the ephemeral conversation context (which is lost after compaction or session end), memory files accumulate learnings, preferences, and project knowledge over time.

Architecture

graph TD
    subgraph "Memory Directory (~/.claude/projects/<project>/memory/)"
        E["MEMORY.md<br/>(index file, max 200 lines)"]
        T1["topic-1.md"]
        T2["topic-2.md"]
        L["logs/YYYY/MM/YYYY-MM-DD.md<br/>(daily logs)"]
    end

    subgraph "System Prompt"
        SP[Memory Prompt Section]
    end

    subgraph "Team Memory"
        TM["Team member shared memory"]
    end

    E --> SP
    T1 --> SP
    TM --> SP

The memory directory follows a standard structure:

• MEMORY.md — the index/entrypoint file, loaded into every system prompt
• Topic files — detailed memory on specific topics, referenced from MEMORY.md
• Daily logs — chronological activity logs (used in KAIROS/assistant mode)

Path Resolution

Memory paths are resolved by getAutoMemPath() in src/memdir/paths.ts with a priority chain:

src/memdir/paths.ts

export const getAutoMemPath = memoize((): string => {
  // 1. Env var override (Cowork/SDK)
  const override = getAutoMemPathOverride() ?? getAutoMemPathSetting()
  if (override) return override

  // 2. Default: ~/.claude/projects/<sanitized-git-root>/memory/
  const projectsDir = join(getMemoryBaseDir(), 'projects')
  return join(projectsDir, sanitizePath(getAutoMemBase()), AUTO_MEM_DIRNAME) + sep
})

The resolution order:

Priority	Source	Use Case
1	CLAUDE_COWORK_MEMORY_PATH_OVERRIDE env var	Cowork space-scoped mount
2	autoMemoryDirectory in settings.json	User custom directory
3	~/.claude/projects/<sanitized-path>/memory/	Default

警告

Security: The settings.json autoMemoryDirectory is only read from trusted sources (policy, local, user settings). Project settings (.claude/settings.json committed to the repo) are explicitly excluded — a malicious repo could set autoMemoryDirectory: "~/.ssh" and gain write access via the filesystem write carve-out.

Git Worktree Awareness

The base path uses findCanonicalGitRoot() so all worktrees of the same repository share one memory directory:

function getAutoMemBase(): string {
  return findCanonicalGitRoot(getProjectRoot()) ?? getProjectRoot()
}

Enable/Disable Logic

Auto-memory can be disabled through multiple mechanisms:

src/memdir/paths.ts

export function isAutoMemoryEnabled(): boolean {
  // 1. Env var override (explicit off)
  if (isEnvTruthy(process.env.CLAUDE_CODE_DISABLE_AUTO_MEMORY)) return false

  // 2. Bare mode (--bare / SIMPLE)
  if (isEnvTruthy(process.env.CLAUDE_CODE_SIMPLE)) return false

  // 3. CCR without persistent storage
  if (isEnvTruthy(process.env.CLAUDE_CODE_REMOTE) &&
      !process.env.CLAUDE_CODE_REMOTE_MEMORY_DIR) return false

  // 4. Settings.json opt-out
  const settings = getInitialSettings()
  if (settings.autoMemoryEnabled !== undefined) return settings.autoMemoryEnabled

  // 5. Default: enabled
  return true
}

MEMORY.md: The Index File

MEMORY.md is the entrypoint that gets loaded into the system prompt. It has strict size limits:

src/memdir/memdir.ts

export const ENTRYPOINT_NAME = 'MEMORY.md'
export const MAX_ENTRYPOINT_LINES = 200
export const MAX_ENTRYPOINT_BYTES = 25_000  // ~125 chars/line at 200 lines

When MEMORY.md exceeds these limits, it’s truncated with a warning:

export function truncateEntrypointContent(raw: string): EntrypointTruncation {
  const contentLines = raw.trim().split('\n')
  const wasLineTruncated = contentLines.length > MAX_ENTRYPOINT_LINES
  const wasByteTruncated = raw.trim().length > MAX_ENTRYPOINT_BYTES

  if (!wasLineTruncated && !wasByteTruncated) {
    return { content: raw.trim(), /* ... */ }
  }

  // Line-truncate first (natural boundary), then byte-truncate
  let truncated = wasLineTruncated
    ? contentLines.slice(0, MAX_ENTRYPOINT_LINES).join('\n')
    : raw.trim()

  if (truncated.length > MAX_ENTRYPOINT_BYTES) {
    const cutAt = truncated.lastIndexOf('\n', MAX_ENTRYPOINT_BYTES)
    truncated = truncated.slice(0, cutAt > 0 ? cutAt : MAX_ENTRYPOINT_BYTES)
  }

  return {
    content: truncated + `\n\n> WARNING: ${ENTRYPOINT_NAME} is ${reason}.
Only part of it was loaded. Keep index entries to one line under ~200 chars;
move detail into topic files.`,
    // ...
  }
}

The design forces MEMORY.md to be an index, not a dump. Each entry should be one line under ~200 chars, pointing to topic files for details.

Memory Types

Memory content is categorized by type, defined in src/memdir/memoryTypes.ts:

// Memory taxonomy (from memoryTypes.ts)
// - User preferences and style
// - Project architecture and conventions
// - Feedback and corrections
// - Reference documentation

The memory prompt includes specific guidance on what to save and what not to save:

import {
  MEMORY_FRONTMATTER_EXAMPLE,
  TRUSTING_RECALL_SECTION,
  TYPES_SECTION_INDIVIDUAL,
  WHAT_NOT_TO_SAVE_SECTION,
  WHEN_TO_ACCESS_SECTION,
} from './memoryTypes.js'

Agent-Specific Memory

Custom agents can declare their own memory scope:

// In agent frontmatter or JSON
memory: 'user' | 'project' | 'local'

Scope	Location	Shared?
user	User-level directory	Across all projects
project	Project-level directory	Within project
local	Local-only	Not shared

When an agent has memory enabled, its system prompt is dynamically appended:

src/tools/AgentTool/loadAgentsDir.ts

getSystemPrompt: () => {
  if (isAutoMemoryEnabled() && memory) {
    return systemPrompt + '\n\n' + loadAgentMemoryPrompt(agentType, memory)
  }
  return systemPrompt
}

Additionally, file tools (Read, Edit, Write) are automatically injected into the agent’s tool set when memory is enabled:

if (isAutoMemoryEnabled() && parsed.memory && tools !== undefined) {
  const toolSet = new Set(tools)
  for (const tool of [FILE_WRITE_TOOL_NAME, FILE_EDIT_TOOL_NAME, FILE_READ_TOOL_NAME]) {
    if (!toolSet.has(tool)) {
      tools = [...tools, tool]
    }
  }
}

Memory Snapshots

The AGENT_MEMORY_SNAPSHOT feature enables project-level memory snapshots — a way to distribute pre-built memory to all users of a project:

src/tools/AgentTool/agentMemorySnapshot.ts

async function initializeAgentMemorySnapshots(agents: CustomAgentDefinition[]): Promise<void> {
  await Promise.all(agents.map(async agent => {
    if (agent.memory !== 'user') return
    const result = await checkAgentMemorySnapshot(agent.agentType, agent.memory)
    switch (result.action) {
      case 'initialize':
        // Copy snapshot to local if no local memory exists
        await initializeFromSnapshot(agent.agentType, agent.memory, result.snapshotTimestamp!)
        break
      case 'prompt-update':
        // Newer snapshot available — flag for user prompt
        agent.pendingSnapshotUpdate = { snapshotTimestamp: result.snapshotTimestamp! }
        break
    }
  }))
}

Team Memory

The TEAMMEM feature enables shared memory across team members:

// src/memdir/memdir.ts — conditional import
const teamMemPaths = feature('TEAMMEM')
  ? require('./teamMemPaths.js')
  : null

Team memory paths (src/memdir/teamMemPaths.ts) and team memory prompts (src/memdir/teamMemPrompts.ts) allow teammates to share a common memory directory, enabling knowledge transfer between agents in the same team.

Extract Mode

The EXTRACT_MEMORIES feature runs a background agent that extracts memories from the conversation:

src/memdir/paths.ts

export function isExtractModeActive(): boolean {
  if (!getFeatureValue_CACHED_MAY_BE_STALE('tengu_passport_quail', false)) return false
  return !getIsNonInteractiveSession() ||
    getFeatureValue_CACHED_MAY_BE_STALE('tengu_slate_thimble', false)
}

The background extraction agent scans conversation turns for memorable information and writes it to the memory directory. When the main agent writes memories itself, the background agent skips those ranges (deduplication via hasMemoryWritesSince).

Daily Log Mode

In KAIROS/assistant sessions, memory takes a different form — daily log files:

src/memdir/paths.ts

export function getAutoMemDailyLogPath(date: Date = new Date()): string {
  const yyyy = date.getFullYear().toString()
  const mm = (date.getMonth() + 1).toString().padStart(2, '0')
  const dd = date.getDate().toString().padStart(2, '0')
  return join(getAutoMemPath(), 'logs', yyyy, mm, `${yyyy}-${mm}-${dd}.md`)
}

Rather than maintaining MEMORY.md as a live index, the agent appends to date-named log files as it works. A separate nightly /dream skill distills these logs into topic files and MEMORY.md.

Path Validation

Memory paths undergo strict security validation:

// src/memdir/paths.ts — validateMemoryPath
function validateMemoryPath(raw: string | undefined, expandTilde: boolean): string | undefined {
  // Reject: relative paths, root/near-root, Windows drive-root,
  //         UNC paths, null bytes
  if (
    !isAbsolute(normalized) ||
    normalized.length < 3 ||
    /^[A-Za-z]:$/.test(normalized) ||
    normalized.startsWith('\\\\') ||
    normalized.startsWith('//') ||
    normalized.includes('\0')
  ) {
    return undefined
  }
  return (normalized + sep).normalize('NFC')
}

The isAutoMemPath() function is used by the filesystem permission system to grant write access to memory files without explicit user permission — a carve-out that makes memory writes seamless but requires careful path validation to prevent abuse.