diff --git a/src/cli/commands/capture.ts b/src/cli/commands/capture.ts
new file mode 100644
index 0000000..3d72b14
--- /dev/null
+++ b/src/cli/commands/capture.ts
@@ -0,0 +1,176 @@
+import { Command } from 'commander';
+import chalk from 'chalk';
+import { captureConversation, captureText, getCaptureConfig, setCaptureConfig, CaptureMode } from '../../core/capture';
+
+const VALID_MODES: CaptureMode[] = ['always', 'manual', 'decisions', 'off'];
+
+export const captureCommand = new Command('capture')
+  .description('Capture text or conversation as a memory node')
+  .argument('[text]', 'Text to capture (or pipe via stdin)')
+  .option('--tags <tags>', 'Comma-separated tags')
+  .option('--source <source>', 'Source identifier', 'manual')
+  .action(async (text: string | undefined, opts) => {
+    let content = text;
+
+    // Read from stdin if no text provided
+    if (!content) {
+      content = await readStdin();
+    }
+
+    if (!content || content.trim().length === 0) {
+      console.error(chalk.red('No text provided. Pass text as argument or pipe via stdin.'));
+      process.exit(1);
+    }
+
+    const tags = opts.tags ? opts.tags.split(',').map((t: string) => t.trim()) : [];
+
+    const result = await captureText(content, { tags, source: opts.source });
+
+    if (!result.captured) {
+      console.log(chalk.yellow(`⚠ Not captured: ${result.reason}`));
+      return;
+    }
+
+    console.log(chalk.green(`✓ Memory ${result.action}`));
+    if (result.node) {
+      console.log(`  ID:    ${chalk.cyan(result.node.id)}`);
+      console.log(`  Title: ${result.node.title}`);
+      if (result.node.tags.length) console.log(`  Tags:  ${result.node.tags.join(', ')}`);
+    }
+    if (result.reason) {
+      console.log(`  ${chalk.dim(result.reason)}`);
+    }
+  });
+
+export const captureHookCommand = new Command('capture-hook')
+  .description('Hook handler for Claude Code auto-capture (receives JSON on stdin)')
+  .option('--session <id>', 'Session ID')
+  .action(async (opts) => {
+    const input = await readStdin();
+
+    if (!input) {
+      // Silent exit - hook may be called with empty input
+      process.exit(0);
+    }
+
+    let data: { conversation?: string; files_changed?: string[]; session_id?: string };
+    try {
+      data = JSON.parse(input);
+    } catch {
+      // Not JSON, treat as plain conversation text
+      data = { conversation: input };
+    }
+
+    const conversation = data.conversation;
+    if (!conversation) {
+      process.exit(0);
+    }
+
+    const result = await captureConversation({
+      conversation,
+      sessionId: data.session_id || opts.session,
+      filesChanged: data.files_changed,
+      source: 'claude-code',
+    });
+
+    // Output result as JSON for hook system
+    console.log(JSON.stringify({
+      captured: result.captured,
+      action: result.action,
+      nodeId: result.node?.id,
+      reason: result.reason,
+    }));
+  });
+
+export const configCommand = new Command('config')
+  .description('Manage capture configuration')
+  .argument('<action>', 'Action: get, set, or list')
+  .argument('[key]', 'Config key (for get/set)')
+  .argument('[value]', 'Config value (for set)')
+  .action(async (action: string, key?: string, value?: string) => {
+    const config = getCaptureConfig();
+
+    if (action === 'list') {
+      console.log(chalk.bold('Capture Configuration:'));
+      console.log(`  mode:              ${chalk.cyan(config.mode)}`);
+      console.log(`  minLength:         ${config.minLength}`);
+      console.log(`  excludePatterns:   ${config.excludePatterns.length ? config.excludePatterns.join(', ') : chalk.dim('(none)')}`);
+      console.log(`  autoTag:           ${config.autoTag}`);
+      console.log(`  linkRelated:       ${config.linkRelated}`);
+      console.log(`  similarityThreshold: ${config.similarityThreshold}`);
+      console.log(`  mergeThreshold:    ${config.mergeThreshold}`);
+      return;
+    }
+
+    if (action === 'get') {
+      if (!key) {
+        console.error(chalk.red('Key required for get'));
+        process.exit(1);
+      }
+      if (!(key in config)) {
+        console.error(chalk.red(`Unknown config key: ${key}`));
+        process.exit(1);
+      }
+      const val = config[key as keyof typeof config];
+      console.log(Array.isArray(val) ? val.join(', ') : String(val));
+      return;
+    }
+
+    if (action === 'set') {
+      if (!key || value === undefined) {
+        console.error(chalk.red('Key and value required for set'));
+        process.exit(1);
+      }
+
+      let parsedValue: any = value;
+
+      // Parse value based on key type
+      if (key === 'mode') {
+        if (!VALID_MODES.includes(value as CaptureMode)) {
+          console.error(chalk.red(`Invalid mode. Must be one of: ${VALID_MODES.join(', ')}`));
+          process.exit(1);
+        }
+        parsedValue = value;
+      } else if (key === 'minLength' || key === 'similarityThreshold' || key === 'mergeThreshold') {
+        parsedValue = parseFloat(value);
+        if (isNaN(parsedValue)) {
+          console.error(chalk.red('Value must be a number'));
+          process.exit(1);
+        }
+      } else if (key === 'autoTag' || key === 'linkRelated') {
+        parsedValue = value === 'true' || value === '1';
+      } else if (key === 'excludePatterns') {
+        parsedValue = value.split(',').map(s => s.trim()).filter(Boolean);
+      } else {
+        console.error(chalk.red(`Unknown config key: ${key}`));
+        process.exit(1);
+      }
+
+      setCaptureConfig({ [key]: parsedValue });
+      console.log(chalk.green(`✓ Set ${key} = ${JSON.stringify(parsedValue)}`));
+      return;
+    }
+
+    console.error(chalk.red('Invalid action. Use: get, set, or list'));
+    process.exit(1);
+  });
+
+async function readStdin(): Promise<string> {
+  return new Promise((resolve) => {
+    // Check if stdin has data (non-TTY mode)
+    if (process.stdin.isTTY) {
+      resolve('');
+      return;
+    }
+
+    let data = '';
+    process.stdin.setEncoding('utf8');
+    process.stdin.on('data', (chunk) => { data += chunk; });
+    process.stdin.on('end', () => { resolve(data.trim()); });
+
+    // Timeout after 100ms if no data
+    setTimeout(() => {
+      if (!data) resolve('');
+    }, 100);
+  });
+}
diff --git a/src/cli/index.ts b/src/cli/index.ts
index 0a1f253..d08af04 100644
--- a/src/cli/index.ts
+++ b/src/cli/index.ts
@@ -11,6 +11,7 @@ import { graphCommand } from './commands/graph';
 import { serveCommand } from './commands/serve';
 import { decayCommand } from './commands/decay';
 import { childrenCommand } from './commands/children';
+import { captureCommand, captureHookCommand, configCommand } from './commands/capture';
 import { closeDb } from '../core/db';
 
 const program = new Command();
@@ -31,6 +32,9 @@ program.addCommand(graphCommand);
 program.addCommand(serveCommand);
 program.addCommand(decayCommand);
 program.addCommand(childrenCommand);
+program.addCommand(captureCommand);
+program.addCommand(captureHookCommand);
+program.addCommand(configCommand);
 
 program.hook('postAction', () => {
   closeDb();
diff --git a/src/core/capture/config.ts b/src/core/capture/config.ts
new file mode 100644
index 0000000..6b21887
--- /dev/null
+++ b/src/core/capture/config.ts
@@ -0,0 +1,68 @@
+import { getDb } from '../db';
+
+export type CaptureMode = 'always' | 'manual' | 'decisions' | 'off';
+
+export interface CaptureConfig {
+  mode: CaptureMode;
+  minLength: number;
+  excludePatterns: string[];
+  autoTag: boolean;
+  linkRelated: boolean;
+  similarityThreshold: number;
+  mergeThreshold: number;
+}
+
+const DEFAULT_CONFIG: CaptureConfig = {
+  mode: 'always',
+  minLength: 100,
+  excludePatterns: [],
+  autoTag: true,
+  linkRelated: true,
+  similarityThreshold: 0.75,
+  mergeThreshold: 0.90,
+};
+
+function ensureConfigTable(): void {
+  const db = getDb();
+  db.exec(`
+    CREATE TABLE IF NOT EXISTS system_config (
+      key TEXT PRIMARY KEY,
+      value TEXT NOT NULL,
+      updated_at INTEGER NOT NULL
+    )
+  `);
+}
+
+export function getCaptureConfig(): CaptureConfig {
+  ensureConfigTable();
+  const db = getDb();
+  const row = db.prepare('SELECT value FROM system_config WHERE key = ?').get('capture') as { value: string } | undefined;
+  if (!row) return DEFAULT_CONFIG;
+  try {
+    return { ...DEFAULT_CONFIG, ...JSON.parse(row.value) };
+  } catch {
+    return DEFAULT_CONFIG;
+  }
+}
+
+export function setCaptureConfig(updates: Partial<CaptureConfig>): CaptureConfig {
+  ensureConfigTable();
+  const db = getDb();
+  const current = getCaptureConfig();
+  const updated = { ...current, ...updates };
+
+  db.prepare(`
+    INSERT INTO system_config (key, value, updated_at) VALUES (?, ?, ?)
+    ON CONFLICT(key) DO UPDATE SET value = excluded.value, updated_at = excluded.updated_at
+  `).run('capture', JSON.stringify(updated), Date.now());
+
+  return updated;
+}
+
+export function getConfigValue<K extends keyof CaptureConfig>(key: K): CaptureConfig[K] {
+  return getCaptureConfig()[key];
+}
+
+export function setConfigValue<K extends keyof CaptureConfig>(key: K, value: CaptureConfig[K]): void {
+  setCaptureConfig({ [key]: value } as Partial<CaptureConfig>);
+}
diff --git a/src/core/capture/dedupe.ts b/src/core/capture/dedupe.ts
new file mode 100644
index 0000000..6f918e7
--- /dev/null
+++ b/src/core/capture/dedupe.ts
@@ -0,0 +1,113 @@
+import { listNodes, addEdge, updateNode } from '../store';
+import { getEmbedding } from '../search/ollama';
+import { cosineSimilarity } from '../search/vector';
+import { Node } from '../../types';
+import { getCaptureConfig } from './config';
+
+export interface SimilarNode {
+  node: Node;
+  similarity: number;
+}
+
+export interface DedupeResult {
+  action: 'create' | 'merge' | 'link';
+  existingNode?: Node;
+  similarity?: number;
+}
+
+export async function findSimilarNodes(
+  text: string,
+  limit: number = 5
+): Promise<SimilarNode[]> {
+  const embedding = await getEmbedding(text);
+  if (!embedding) return [];
+
+  const nodes = listNodes({ includeStale: false });
+  const withEmbeddings = nodes.filter(n => n.embedding && n.embedding.length > 0);
+
+  const scored: SimilarNode[] = [];
+  for (const node of withEmbeddings) {
+    const similarity = cosineSimilarity(embedding, node.embedding!);
+    if (similarity > 0.5) {
+      scored.push({ node, similarity });
+    }
+  }
+
+  return scored
+    .sort((a, b) => b.similarity - a.similarity)
+    .slice(0, limit);
+}
+
+export async function checkDuplicate(
+  summary: string,
+  content: string
+): Promise<DedupeResult> {
+  const config = getCaptureConfig();
+  const textToCompare = `${summary} ${content}`;
+
+  const similar = await findSimilarNodes(textToCompare, 1);
+
+  if (similar.length === 0) {
+    return { action: 'create' };
+  }
+
+  const { node, similarity } = similar[0];
+
+  if (similarity >= config.mergeThreshold) {
+    return {
+      action: 'merge',
+      existingNode: node,
+      similarity,
+    };
+  }
+
+  if (similarity >= config.similarityThreshold) {
+    return {
+      action: 'link',
+      existingNode: node,
+      similarity,
+    };
+  }
+
+  return { action: 'create' };
+}
+
+export async function mergeIntoNode(
+  existingId: string,
+  newSummary: string,
+  newContent: string,
+  newTags: string[]
+): Promise<Node | null> {
+  const existing = listNodes({ includeStale: false }).find(n => n.id === existingId);
+  if (!existing) return null;
+
+  // Append new content with timestamp
+  const timestamp = new Date().toISOString().slice(0, 10);
+  const mergedContent = existing.content
+    ? `${existing.content}\n\n---\n[${timestamp}]\n${newContent}`
+    : newContent;
+
+  // Merge tags (dedupe)
+  const mergedTags = [...new Set([...existing.tags, ...newTags])];
+
+  // Update the existing node
+  return updateNode(existingId, {
+    content: mergedContent,
+    tags: mergedTags,
+    metadata: {
+      ...existing.metadata,
+      lastMergedAt: Date.now(),
+      mergeCount: (existing.metadata.mergeCount || 0) + 1,
+    },
+  });
+}
+
+export async function linkRelatedNode(
+  newNodeId: string,
+  existingNodeId: string
+): Promise<void> {
+  addEdge(newNodeId, existingNodeId, 'relates_to', {
+    reason: 'auto-capture-similarity',
+    linkedAt: Date.now(),
+  });
+}
diff --git a/src/core/capture/index.ts b/src/core/capture/index.ts
new file mode 100644
index 0000000..8dcc0cc
--- /dev/null
+++ b/src/core/capture/index.ts
@@ -0,0 +1,192 @@
+import { addNode } from '../store';
+import { getCaptureConfig, CaptureConfig } from './config';
+import { extractMemoryData, shouldCapture, ExtractedMemory } from './summarize';
+import { checkDuplicate, mergeIntoNode, linkRelatedNode } from './dedupe';
+import { Node } from '../../types';
+
+export { getCaptureConfig, setCaptureConfig, CaptureMode, CaptureConfig } from './config';
+export { extractMemoryData, shouldCapture, ExtractedMemory } from './summarize';
+export { findSimilarNodes, checkDuplicate, mergeIntoNode } from './dedupe';
+
+export interface CaptureInput {
+  conversation: string;
+  sessionId?: string;
+  filesChanged?: string[];
+  source?: string;
+}
+
+export interface CaptureResult {
+  captured: boolean;
+  action: 'created' | 'merged' | 'linked' | 'skipped';
+  node?: Node;
+  reason?: string;
+}
+
+export async function captureConversation(input: CaptureInput): Promise<CaptureResult> {
+  const config = getCaptureConfig();
+
+  // Check if capture is enabled
+  if (config.mode === 'off') {
+    return { captured: false, action: 'skipped', reason: 'capture disabled' };
+  }
+
+  // Check minimum length
+  if (!shouldCapture(input.conversation, config.minLength)) {
+    return { captured: false, action: 'skipped', reason: 'conversation too short or trivial' };
+  }
+
+  // Check exclude patterns
+  for (const pattern of config.excludePatterns) {
+    try {
+      if (new RegExp(pattern, 'i').test(input.conversation)) {
+        return { captured: false, action: 'skipped', reason: `matched exclude pattern: ${pattern}` };
+      }
+    } catch {
+      // Invalid regex, skip
+    }
+  }
+
+  // Extract memory data using Ollama
+  const extracted = await extractMemoryData(input.conversation);
+  if (!extracted) {
+    return { captured: false, action: 'skipped', reason: 'failed to extract memory data' };
+  }
+
+  // For "decisions" mode, only capture if decisions were found
+  if (config.mode === 'decisions' && extracted.decisions.length === 0) {
+    return { captured: false, action: 'skipped', reason: 'no decisions found (decisions mode)' };
+  }
+
+  // Build content
+  const contentParts: string[] = [extracted.summary];
+
+  if (extracted.decisions.length > 0) {
+    contentParts.push('\n## Decisions');
+    for (const d of extracted.decisions) {
+      contentParts.push(`- ${d}`);
+    }
+  }
+
+  if (extracted.filesDiscussed.length > 0 || input.filesChanged?.length) {
+    const files = [...new Set([...extracted.filesDiscussed, ...(input.filesChanged || [])])];
+    contentParts.push('\n## Files');
+    for (const f of files) {
+      contentParts.push(`- ${f}`);
+    }
+  }
+
+  const content = contentParts.join('\n');
+
+  // Check for duplicates
+  const dedupeResult = await checkDuplicate(extracted.summary, content);
+
+  // Build tags
+  const tags = ['auto-capture'];
+  if (config.autoTag && extracted.topics.length > 0) {
+    tags.push(...extracted.topics);
+  }
+  if (input.source) {
+    tags.push(`source:${input.source}`);
+  }
+
+  if (dedupeResult.action === 'merge' && dedupeResult.existingNode) {
+    // Merge into existing node
+    const merged = await mergeIntoNode(
+      dedupeResult.existingNode.id,
+      extracted.summary,
+      content,
+      tags
+    );
+    return {
+      captured: true,
+      action: 'merged',
+      node: merged || undefined,
+      reason: `merged with existing node (similarity: ${(dedupeResult.similarity! * 100).toFixed(1)}%)`,
+    };
+  }
+
+  // Create new node
+  const node = await addNode({
+    kind: 'memory',
+    title: extracted.summary.slice(0, 100),
+    content,
+    tags,
+    status: 'active',
+    metadata: {
+      sessionId: input.sessionId,
+      filesChanged: input.filesChanged,
+      source: input.source || 'claude-code',
+      capturedAt: Date.now(),
+      decisions: extracted.decisions,
+    },
+  });
+
+  // Link to related node if found
+  if (dedupeResult.action === 'link' && dedupeResult.existingNode && config.linkRelated) {
+    await linkRelatedNode(node.id, dedupeResult.existingNode.id);
+    return {
+      captured: true,
+      action: 'linked',
+      node,
+      reason: `linked to related node (similarity: ${(dedupeResult.similarity! * 100).toFixed(1)}%)`,
+    };
+  }
+
+  return {
+    captured: true,
+    action: 'created',
+    node,
+  };
+}
+
+export async function captureText(
+  text: string,
+  options: { tags?: string[]; source?: string } = {}
+): Promise<CaptureResult> {
+  const config = getCaptureConfig();
+
+  if (config.mode === 'off') {
+    return { captured: false, action: 'skipped', reason: 'capture disabled' };
+  }
+
+  // Simple text capture - no summarization needed
+  const dedupeResult = await checkDuplicate(text, text);
+
+  const tags = ['manual-capture', ...(options.tags || [])];
+  if (options.source) {
+    tags.push(`source:${options.source}`);
+  }
+
+  if (dedupeResult.action === 'merge' && dedupeResult.existingNode) {
+    const merged = await mergeIntoNode(
+      dedupeResult.existingNode.id,
+      text.slice(0, 100),
+      text,
+      tags
+    );
+    return {
+      captured: true,
+      action: 'merged',
+      node: merged || undefined,
+    };
+  }
+
+  const node = await addNode({
+    kind: 'memory',
+    title: text.slice(0, 100),
+    content: text,
+    tags,
+    status: 'active',
+    metadata: {
+      source: options.source || 'manual',
+      capturedAt: Date.now(),
+    },
+  });
+
+  if (dedupeResult.action === 'link' && dedupeResult.existingNode && config.linkRelated) {
+    await linkRelatedNode(node.id, dedupeResult.existingNode.id);
+    return { captured: true, action: 'linked', node };
+  }
+
+  return { captured: true, action: 'created', node };
+}
diff --git a/src/core/capture/summarize.ts b/src/core/capture/summarize.ts
new file mode 100644
index 0000000..8bd1b26
--- /dev/null
+++ b/src/core/capture/summarize.ts
@@ -0,0 +1,160 @@
+import { generate, isGenAvailable } from '../search/ollamaGen';
+
+export interface ExtractedMemory {
+  summary: string;
+  topics: string[];
+  decisions: string[];
+  filesDiscussed: string[];
+}
+
+const SUMMARIZE_PROMPT = `Summarize this Claude Code conversation in 1-2 sentences.
+Focus on: what was accomplished, decisions made, problems solved.
+Do NOT include greetings or meta-discussion.
+
+Conversation:
+{conversation}
+
+Summary:`;
+
+const EXTRACT_PROMPT = `Extract from this conversation:
+1. Main topics (as tags, lowercase, hyphenated, max 5)
+2. Decisions made (if any, max 3)
+3. Code files discussed or modified (if any)
+
+Conversation:
+{conversation}
+
+Output as JSON only, no explanation:
+{"topics": [], "decisions": [], "files": []}`;
+
+export async function summarizeConversation(conversation: string): Promise<string | null> {
+  if (!(await isGenAvailable())) return null;
+
+  const prompt = SUMMARIZE_PROMPT.replace('{conversation}', conversation);
+  return generate(prompt);
+}
+
+export async function extractMemoryData(conversation: string): Promise<ExtractedMemory | null> {
+  const available = await isGenAvailable();
+
+  // Get summary
+  const summary = available
+    ? await summarizeConversation(conversation)
+    : createFallbackSummary(conversation);
+
+  if (!summary) return null;
+
+  // Extract structured data
+  let topics: string[] = [];
+  let decisions: string[] = [];
+  let filesDiscussed: string[] = [];
+
+  if (available) {
+    const extractPrompt = EXTRACT_PROMPT.replace('{conversation}', conversation);
+    const extracted = await generate(extractPrompt);
+
+    if (extracted) {
+      try {
+        // Find JSON in response (handle cases where model adds explanation)
+        const jsonMatch = extracted.match(/\{[\s\S]*\}/);
+        if (jsonMatch) {
+          const data = JSON.parse(jsonMatch[0]);
+          topics = Array.isArray(data.topics) ? data.topics.slice(0, 5) : [];
+          decisions = Array.isArray(data.decisions) ? data.decisions.slice(0, 3) : [];
+          filesDiscussed = Array.isArray(data.files) ? data.files : [];
+        }
+      } catch {
+        // Fall back to basic extraction
+        topics = extractTopicsBasic(conversation);
+        filesDiscussed = extractFilesBasic(conversation);
+      }
+    }
+  } else {
+    // Basic extraction without AI
+    topics = extractTopicsBasic(conversation);
+    filesDiscussed = extractFilesBasic(conversation);
+  }
+
+  return {
+    summary,
+    topics: sanitizeTags(topics),
+    decisions,
+    filesDiscussed,
+  };
+}
+
+function createFallbackSummary(conversation: string): string {
+  // Take first meaningful line as summary
+  const lines = conversation.split('\n').filter(l => l.trim().length > 20);
+  if (lines.length === 0) return 'Conversation captured';
+
+  const first = lines[0].trim();
+  return first.length > 150 ? first.slice(0, 147) + '...' : first;
+}
+
+function extractTopicsBasic(conversation: string): string[] {
+  const topics: string[] = [];
+  const lower = conversation.toLowerCase();
+
+  // Common programming topics
+  const keywords = [
+    'typescript', 'javascript', 'python', 'rust', 'go',
+    'react', 'vue', 'angular', 'node', 'express',
+    'database', 'sql', 'api', 'auth', 'authentication',
+    'bug', 'fix', 'error', 'refactor', 'test', 'deploy',
+    'git', 'docker', 'kubernetes', 'aws', 'cloud',
+  ];
+
+  for (const kw of keywords) {
+    if (lower.includes(kw) && topics.length < 5) {
+      topics.push(kw);
+    }
+  }
+
+  return topics;
+}
+
+function extractFilesBasic(conversation: string): string[] {
+  const files: string[] = [];
+
+  // Match file paths
+  const filePatterns = [
+    /[\w\-\/]+\.(ts|js|tsx|jsx|py|rs|go|md|json|yaml|yml|toml|sql)/gi,
+    /src\/[\w\-\/]+/gi,
+  ];
+
+  for (const pattern of filePatterns) {
+    const matches = conversation.match(pattern);
+    if (matches) {
+      for (const m of matches) {
+        if (!files.includes(m) && files.length < 10) {
+          files.push(m);
+        }
+      }
+    }
+  }
+
+  return files;
+}
+
+function sanitizeTags(tags: string[]): string[] {
+  return tags
+    .map(t => t.toLowerCase().trim().replace(/\s+/g, '-').replace(/[^a-z0-9\-]/g, ''))
+    .filter(t => t.length > 0 && t.length < 30);
+}
+
+export function shouldCapture(conversation: string, minLength: number): boolean {
+  // Skip very short conversations
+  if (conversation.length < minLength) return false;
+
+  // Skip if mostly greetings/pleasantries
+  const lower = conversation.toLowerCase();
+  const greetings = ['hello', 'hi ', 'hey', 'thanks', 'thank you', 'goodbye', 'bye'];
+  const greetingCount = greetings.filter(g => lower.includes(g)).length;
+
+  // If more than half the "content" is greetings, skip
+  const words = conversation.split(/\s+/).length;
+  if (words < 20 && greetingCount > 2) return false;
+
+  return true;
+}
diff --git a/src/core/summary.ts b/src/core/summary.ts
new file mode 100644
index 0000000..7e46f3a
--- /dev/null
+++ b/src/core/summary.ts
@@ -0,0 +1,165 @@
+import { getDb } from './db';
+import { isGenAvailable, generate } from './search/ollamaGen';
+
+export interface SummaryData {
+  generatedAt: number;
+  overview: string;
+  components: { title: string; status?: string; summary?: string }[];
+  decisions: { title: string; status?: string }[];
+  tasks: { todo: string[]; in_progress: string[]; done: string[] };
+  memories: { byTag: Record<string, string[]> };
+  stats: { total: number; stale: number; orphans: number; edges: number };
+}
+
+// Ensure cache table exists
+function ensureCacheTable(): void {
+  const db = getDb();
+  db.exec(`
+    CREATE TABLE IF NOT EXISTS system_cache (
+      key TEXT PRIMARY KEY,
+      value TEXT NOT NULL,
+      updated_at INTEGER NOT NULL
+    )
+  `);
+}
+
+export function getCachedSummary(): SummaryData | null {
+  ensureCacheTable();
+  const db = getDb();
+  const row = db.prepare('SELECT value FROM system_cache WHERE key = ?').get('summary') as { value: string } | undefined;
+  if (!row) return null;
+  try {
+    return JSON.parse(row.value);
+  } catch {
+    return null;
+  }
+}
+
+export function cacheSummary(data: SummaryData): void {
+  ensureCacheTable();
+  const db = getDb();
+  db.prepare(`
+    INSERT INTO system_cache (key, value, updated_at) VALUES (?, ?, ?)
+    ON CONFLICT(key) DO UPDATE SET value = excluded.value, updated_at = excluded.updated_at
+  `).run('summary', JSON.stringify(data), Date.now());
+}
+
+export async function generateSummary(): Promise<SummaryData> {
+  const db = getDb();
+  const now = Date.now();
+
+  // Gather all active nodes
+  const rows = db.prepare(`
+    SELECT id, kind, title, status, tags, metadata, content
+    FROM nodes WHERE is_stale = 0
+    ORDER BY kind, title
+  `).all() as any[];
+
+  // Stats
+  const totalNodes = rows.length;
+  const staleCount = (db.prepare('SELECT COUNT(*) as c FROM nodes WHERE is_stale = 1').get() as any).c;
+  const edgeCount = (db.prepare('SELECT COUNT(*) as c FROM edges').get() as any).c;
+
+  const edgeRows = db.prepare('SELECT from_id, to_id FROM edges').all() as any[];
+  const hasEdge = new Set<string>();
+  for (const e of edgeRows) {
+    hasEdge.add(e.from_id);
+    hasEdge.add(e.to_id);
+  }
+  const orphanCount = rows.filter(r => !hasEdge.has(r.id)).length;
+
+  // Categorize nodes
+  const components: SummaryData['components'] = [];
+  const decisions: SummaryData['decisions'] = [];
+  const tasks: SummaryData['tasks'] = { todo: [], in_progress: [], done: [] };
+  const memoryTags: Record<string, string[]> = {};
+
+  for (const row of rows) {
+    const meta = JSON.parse(row.metadata || '{}');
+
+    if (row.kind === 'component') {
+      components.push({
+        title: row.title,
+        status: row.status,
+        summary: meta.summary || (row.content?.slice(0, 100) + (row.content?.length > 100 ? '...' : '')),
+      });
+    } else if (row.kind === 'decision') {
+      decisions.push({ title: row.title, status: row.status });
+    } else if (row.kind === 'task') {
+      const status = (row.status || 'todo').toLowerCase();
+      if (status === 'done' || status === 'completed') {
+        tasks.done.push(row.title);
+      } else if (status === 'in_progress' || status === 'in-progress') {
+        tasks.in_progress.push(row.title);
+      } else {
+        tasks.todo.push(row.title);
+      }
+    } else if (row.kind === 'memory') {
+      const tags: string[] = JSON.parse(row.tags || '[]');
+      // Group by first tag, or 'untagged'
+      const primaryTag = tags[0] || 'untagged';
+      if (!memoryTags[primaryTag]) memoryTags[primaryTag] = [];
+      memoryTags[primaryTag].push(row.title);
+    }
+  }
+
+  // Sort tags by count descending, limit to top 10
+  const sortedTags = Object.entries(memoryTags)
+    .sort((a, b) => b[1].length - a[1].length)
+    .slice(0, 10);
+  const limitedMemoryTags: Record<string, string[]> = {};
+  for (const [tag, titles] of sortedTags) {
+    // Limit to 5 titles per tag
+    limitedMemoryTags[tag] = titles.slice(0, 5);
+    if (titles.length > 5) {
+      limitedMemoryTags[tag].push(`+${titles.length - 5} more`);
+    }
+  }
+
+  // Generate overview
+  let overview: string;
+  const aiAvailable = await isGenAvailable();
+
+  if (aiAvailable) {
+    const prompt = `Write a single sentence (max 30 words) summarizing this knowledge graph for a developer starting a coding session.
+
+Stats: ${totalNodes} nodes (${components.length} components, ${decisions.length} decisions, ${tasks.todo.length + tasks.in_progress.length + tasks.done.length} tasks, ${totalNodes - components.length - decisions.length - tasks.todo.length - tasks.in_progress.length - tasks.done.length} memories), ${edgeCount} edges.
+
+Components: ${components.map(c => c.title).join(', ')}
+Active tasks: ${[...tasks.todo, ...tasks.in_progress].join(', ') || 'none'}
+
+Output ONLY the summary sentence.`;
+
+    const aiOverview = await generate(prompt);
+    overview = aiOverview || buildFallbackOverview(totalNodes, components.length, decisions.length, tasks, edgeCount);
+  } else {
+    overview = buildFallbackOverview(totalNodes, components.length, decisions.length, tasks, edgeCount);
+  }
+
+  const summary: SummaryData = {
+    generatedAt: now,
+    overview,
+    components,
+    decisions,
+    tasks,
+    memories: { byTag: limitedMemoryTags },
+    stats: { total: totalNodes, stale: staleCount, orphans: orphanCount, edges: edgeCount },
+  };
+
+  cacheSummary(summary);
+  return summary;
+}
+
+function buildFallbackOverview(
+  total: number,
+  components: number,
+  decisions: number,
+  tasks: SummaryData['tasks'],
+  edges: number
+): string {
+  const pending = tasks.todo.length + tasks.in_progress.length;
+  const parts = [`${total} nodes`, `${components} components`, `${decisions} decisions`];
+  if (pending > 0) parts.push(`${pending} open tasks`);
+  parts.push(`${edges} edges`);
+  return parts.join(', ') + '.';
+}
diff --git a/src/mcp/index.ts b/src/mcp/index.ts
index b499fe0..7ff1514 100644
--- a/src/mcp/index.ts
+++ b/src/mcp/index.ts
@@ -385,6 +385,69 @@ server.tool(
   }
 );
 
+// --- memory_capture ---
+import { captureConversation, captureText, getCaptureConfig, setCaptureConfig, CaptureMode } from '../core/capture';
+
+server.tool(
+  'memory_capture',
+  'Capture a conversation or context as a memory node. Uses AI to summarize and extract key information.',
+  {
+    conversation: z.string().describe('The conversation or context to capture'),
+    sessionId: z.string().optional().describe('Session identifier'),
+    filesChanged: z.array(z.string()).optional().describe('List of files that were changed'),
+    source: z.string().optional().describe('Source identifier (default: claude-code)'),
+  },
+  async ({ conversation, sessionId, filesChanged, source }) => {
+    const result = await captureConversation({
+      conversation,
+      sessionId,
+      filesChanged,
+      source: source || 'claude-code',
+    });
+    return { content: [{ type: 'text' as const, text: serialize(result) }] };
+  }
+);
+
+server.tool(
+  'memory_remember',
+  'Remember a piece of text for later. Simpler than memory_capture - for quick notes and facts.',
+  {
+    text: z.string().describe('The text to remember'),
+    tags: z.array(z.string()).optional().describe('Tags to apply'),
+  },
+  async ({ text, tags }) => {
+    const result = await captureText(text, { tags, source: 'remember' });
+    return { content: [{ type: 'text' as const, text: serialize(result) }] };
+  }
+);
+
+server.tool(
+  'memory_capture_config',
+  'Get or set auto-capture configuration',
+  {
+    action: z.enum(['get', 'set']).describe('Action to perform'),
+    mode: z.enum(['always', 'manual', 'decisions', 'off']).optional().describe('Capture mode (for set)'),
+    minLength: z.number().optional().describe('Minimum conversation length (for set)'),
+    autoTag: z.boolean().optional().describe('Auto-generate tags (for set)'),
+    linkRelated: z.boolean().optional().describe('Auto-link related nodes (for set)'),
+  },
+  async ({ action, mode, minLength, autoTag, linkRelated }) => {
+    if (action === 'get') {
+      const config = getCaptureConfig();
+      return { content: [{ type: 'text' as const, text: serialize(config) }] };
+    }
+
+    const updates: Partial<{ mode: CaptureMode; minLength: number; autoTag: boolean; linkRelated: boolean }> = {};
+    if (mode !== undefined) updates.mode = mode;
+    if (minLength !== undefined) updates.minLength = minLength;
+    if (autoTag !== undefined) updates.autoTag = autoTag;
+    if (linkRelated !== undefined) updates.linkRelated = linkRelated;
+
+    const config = setCaptureConfig(updates);
+    return { content: [{ type: 'text' as const, text: serialize({ updated: true, config }) }] };
+  }
+);
+
 // --- memory_prompt ---
 import { interpretAndExecute } from '../core/prompt/interpreter';