Continuity Ultimate
Memory that belongs to your code — not your model.
The Problem
You have explained your architecture to Claude four times this week. Copilot just suggested the exact pattern you rejected last month — and you spent twenty minutes remembering why. Your AI assistant burns tokens rediscovering context, contradicts past decisions, and leaves no trail when things break.
This is not a model problem.
It is a memory problem.
AI assistants forget everything when the session ends. Architectural decisions, constraints, and context vanish. You are not just losing productivity — you are losing control.
The Solution
Continuity creates synthetic memory — a persistent, local-first AI context layer that remembers architectural decisions and feeds them back to Claude, Cursor, or Cline automatically.
Stop re-explaining your codebase.
Start building.
Why This Is Different
Decisions are stored in your repo — not in chat history, not in vendor-managed memory. Your .continuity/ folder travels with your code.
Works with your AI tools today — native MCP integration lets Claude Code, Cursor, and Cline read and write decisions without copy-paste.
Core Capabilities
Persistent Decision Memory — Log architectural decisions once; AI recalls the why across sessions
Code-Linked Context — Bind decisions directly to code locations so AI understands changes in context
Seamless AI Handoff — Auto-generate session context with delta tracking for any MCP-compatible tool
Synthetic Memory Control — Own institutional knowledge locally with zero telemetry and full portability
Continuous Context Sync — Surface decision debt as architecture evolves
What Is Synthetic Memory?
AI sessions are stateless by default. Context evaporates when the tab closes. Some vendors offer "memory," but that state lives outside your control with unclear retention and no auditability.
Synthetic memory inverts this model.
Continuity persists architectural decisions as structured, versioned state inside your repo. You control retention, audit changes over time, and delete state without vendor coordination.
This is owned state for your codebase — not convenience memory for a model.
Privacy & Ownership
100% local. Zero telemetry. Your decisions never leave your machine unless you commit them yourself.
- No cloud dependencies
- No vendor lock-in
- Full git integration
Free vs Pro
Start free with 10 decisions. Verify your email for 5 more. Upgrade for unlimited.
|
Free |
Pro |
| Decisions |
10 (15 w/ email) |
Unlimited |
| MCP Integration |
✓ |
✓ |
| Semantic Search |
✓ |
✓ |
| Code Intelligence |
— |
✓ |
| Team Features |
— |
✓ |
| Pricing |
$0 |
From $15/month |
Pro Plans:
- Monthly: $15/month
- Annual: $139/year (save 23%)
- Lifetime: $399 one-time
Upgrade to Pro →
Get Started
- Install from the VS Code Marketplace
- Setup → Command Palette →
Continuity: Setup MCP
- Log decisions — your AI remembers the rest
Works With
- Claude Code
- Cursor
- Cline / Roo
- GitHub Copilot
- OpenAI Codex
- Google Gemini
- Any MCP-compatible AI tool
What's New
v2.9.9 — Marketplace Release (Mar 2026)
MCP server stability hardening — 7 premature disconnection bugs fixed:
- Grace timer EPIPE fix — stderr pipe errors no longer defeat the stdin grace period;
inGracePeriod flag suppresses Layer 2 during the grace window, preventing silent bypasses when the parent process dies
- Heartbeat reset on SDK pings — server now resets its idle timeout on ANY incoming stdin data, not just tool calls; idle-but-connected sessions (e.g., user reading code for 5+ minutes) no longer trigger premature shutdown
- Auto-path repair for Gemini CLI & Antigravity IDE — extension updates now auto-fix MCP config paths for all 10 supported AI clients, closing a gap where Gemini and Antigravity configs went stale
- Exit code 1 for exceptions —
uncaughtException/unhandledRejection now signal abnormal termination (code 1) so clients expedite server restart instead of treating crashes as clean shutdowns
- Grace timer reduced 30s to 5s — stdio transport can't reconnect, so the old 30-second wait provided no benefit; 5 seconds covers signal ordering races
- Zombie detection tuned — RSS threshold for dormant parent detection lowered from 50MB to 20MB, eliminating false-positive kills on macOS where memory compression pushes healthy idle hosts below 50MB
- 24 new stability tests — keepalive test suite expanded from 20 to 44 tests covering EPIPE suppression, exit codes, timing precision, and combined scenarios
- 951+ architectural decisions tracked across 15 topic categories
v2.9.8 — Marketplace Release (Mar 2026)
Stability, security hardening, and test infrastructure overhaul:
- 30-second keepalive grace timer — MCP server no longer dies instantly on stdin close; tolerates brief client reconnections during IDE restarts and VS Code window reloads, with automatic shutdown if no activity resumes
- 6 sprint security & reliability fixes — embedding cache invalidation on decision update/delete, XSS prevention via consistent
escapeHtml() across all view providers, prompt injection defense with XML delimiter tags and LLM response validation, runtime enum validation for MCP tool parameters, and pre-publish safety gates for marketplace builds
- TOCTOU race condition fix — stale lock detection now uses PID verification to prevent two processes from simultaneously acquiring the same lock
- Free tier enforcement hardened — wired decision limit checks in MCP ServerContext that were left as no-op stubs during modularization
- 208 new tests — enforced 60% coverage threshold in CI; test suite now covers security, licensing, and MCP tool handlers
- 947+ architectural decisions tracked across 15 topic categories
v2.9.3 — Marketplace Release (Mar 2026)
MCP permission automation and Claude Code compatibility improvements:
- MCP wildcard permission auto-setup — extension and setup wizard now auto-configure
mcp__continuity__* wildcard permission in Claude Code settings, replacing individual tool permissions for seamless MCP tool access
- Claude Code .mcp.json compatibility — project-level MCP config now includes required
type: "stdio" field per current Claude Code schema validation
- 941+ architectural decisions tracked across 15 topic categories
v2.9.2 — Marketplace Release (Feb 2026)
Major reliability and context intelligence update — the biggest overhaul since v2.8.2:
- Handoff essentials injection fix — repaired CLAUDE.md injection pipeline that was failing to write project context on session start; AI tools now reliably receive recent decisions, active blockers, and version info at startup
- 6-layer zombie MCP prevention — comprehensive overhaul eliminates orphaned MCP server processes: heartbeat monitoring, stdin-close detection, version-mismatch killing, dormant parent detection, periodic 15-minute sweep, and crash-loop-aware cleanup
- Real-time decision logging — improved logging pipeline captures decisions as they happen with better conflict detection and duplicate filtering
- Cross-workspace MCP auto-migration — config auto-update now migrates ALL workspace entries on extension update, eliminating stale config files that caused zombie accumulation
- Session date tracking — session metadata now includes accurate timestamps for better cross-session handoff continuity
- Engineering guardrails — auto-generated CLAUDE.md now includes mandatory decision-logging rules, ensuring AI assistants log decisions immediately after each change
- 939+ architectural decisions tracked across 15 topic categories
v2.8.2 — Emergency Stability Fix (Feb 2026)
Extension host crash-loop protection and test infrastructure fixes:
- Crash-loop guard — detects rapid extension host restarts (3+ in 5 minutes) and disables aggressive process cleanup that was causing cascading crashes
- Windows process verification —
killPreviousExtensionHost now verifies process identity via tasklist before force-killing, preventing PID-recycling accidents that could kill unrelated VS Code processes
- Orphan cleanup safety — zombie MCP process cleanup is automatically skipped during detected crash loops to break the restart cascade
- Test config fix — Jest
testMatch pattern no longer uses <rootDir> expansion (broken on MINGW/WSL due to path separator mismatch); 46 SecureBank unit tests now discoverable
- Integration runner fix —
run-all.js unit step uses --testPathPattern regex instead of positional path argument, fixing cross-platform test discovery
- 924+ architectural decisions tracked across 15 topic categories
v2.8.1 — Marketplace Release (Feb 2026)
Test suite hardening, multi-agent polish, and persistent project instructions:
- Test suite cleanup — unskipped 9 stale tests, repaired 2 failing suites (tool-profiles counts + SemanticSearch glob mock). 1,140 tests passing, 93 skipped, 0 failures across 45 suites
- Persistent Project Instructions —
.continuity/INSTRUCTIONS.md with dual delivery (MCP + system-prompt injection) now auto-seeds with starter template on first use
read_instructions / update_instructions MCP tools — read and modify project instructions programmatically from any AI toolcontinuity://instructions resource — direct MCP resource access for AI tools that prefer resource-based context loading- Multi-agent instruction sync — CLAUDE.md, AGENTS.md, .cursorrules, copilot-instructions.md, and GEMINI.md refreshed with latest project context and decision history
- MCP server v2.7.5-beta — improved session handoff, context delivery, and instructions integration
- 912+ architectural decisions tracked across 15 topic categories
v2.8.0 (Feb 2026)
Stability, documentation, and multi-agent polish building on the v2.7.x foundation:
- Cross-Platform MCP Portability — portable node paths, WORKSPACE_ROOT auto-repair, Copilot config repair
- Cross-Session Handoff — SESSION_NOTES.md tracked in git for persistent AI context
- Decision Hygiene — post-commit enforcement hooks, database cleanup, build isolation
v2.7.7 (Feb 2026)
All v2.7.4–v2.7.6 features consolidated into a single stable release:
- Persistent Project Instructions, Cross-Platform MCP Portability, Cross-Session Handoff, Decision Hygiene
Persistent Project Instructions (v2.7.6)
.continuity/INSTRUCTIONS.md — a new persistent instructions file that works across all AI tools via MCP. Write your workflow rules, coding conventions, and project-specific notes once; every MCP-connected AI (Claude, Cursor, Copilot, Cline, Gemini, Codex) picks them up automatically.read_instructions tool — reads project instructions on demand, auto-creates a starter template if the file doesn't exist yetupdate_instructions tool — append or replace instructions programmatically (e.g., "remember to always use bun instead of npm")continuity://instructions resource — direct MCP resource access for AI tools that prefer resource-based context loading- System-prompt delivery — instructions are automatically injected into CLAUDE.md, .cursorrules, copilot-instructions.md, GEMINI.md, and AGENTS.md so AI tools see them at startup even without MCP
get_quick_context integration — project instructions are now included in the quick context response when real content existsload_context_section integration — instructions is now a loadable section alongside decisions, git, session_notes, etc.
- Portable MCP configs —
getNodeCommand() now returns "node" instead of resolving the full binary path. Full paths broke when switching between machines, NVM versions, or Node reinstalls. All MCP clients (Copilot, Cursor, Claude Code, Cline) resolve from PATH natively.
- WORKSPACE_ROOT auto-repair —
autoRepairMCPConfigs() now detects and fixes stale workspace environment variables, not just command paths. Switching machines no longer leaves broken MCP configs behind.
- GitHub Copilot config repair —
.vscode/mcp.json is now included in the auto-repair loop alongside Cursor, Cline, and Claude Code configs.
- Machine-specific config isolation —
.vscode/mcp.json is now gitignored by default since it contains machine-specific paths. Each machine generates its own via Setup MCP or auto-repair.
- Cross-session handoff —
SESSION_NOTES.md is now tracked in git, enabling persistent AI context across sessions and machines.
- Repository hygiene — cleaned up 14 stale branches and refreshed all multi-agent instruction files.
Decision Hygiene & Enforcement (v2.7.3)
- Post-commit decision enforcement — new Claude Code hook automatically reminds AI agents to log decisions after every
git commit, eliminating silent decision gaps between sessions
- Decision database cleanup — removed outdated session-specific entries from
decisions.json, keeping the decision log focused on architectural choices that matter
- Documentation sync — updated decision counts, topic indexes, and agent instruction files across CLAUDE.md, AGENTS.md, .cursorrules, copilot-instructions.md, and GEMINI.md to reflect current project state
- Build artifact isolation — removed
continuity_windows_test_UNIT subproject reference to prevent test folder pollution in production builds
- 912+ total decisions tracked across 15 topic categories
Documentation & Configuration Refresh (v2.7.2)
- Refreshed multi-agent instruction files — CLAUDE.md, AGENTS.md, .cursorrules, copilot-instructions.md, and GEMINI.md updated with latest project context, recent decisions, and v2.7 feature documentation
- MCP config auto-cleanup — removed outdated server configurations from
.vscode/mcp.json to prevent stale path references
- Cross-platform path normalization — enhanced MCP server configuration for improved compatibility across macOS, Linux, and Windows environments
- Decision coverage enforcement — backfilled 18+ missing commit decisions and added post-commit hooks to ensure continuous decision logging
Build & Reliability Hardening (v2.7.1)
- Build isolation — test folders are now excluded from TypeScript compilation and webpack bundling, eliminating 52 TS6059 errors that could leak into production builds
- VSIX size enforcement —
.vscodeignore hardened to prevent test artifacts from bloating the marketplace package (keeps VSIX under 10 MB)
- Decision enforcement hook — new Claude Code PostToolUse hook automatically reminds AI to log decisions after every
git commit, closing the gap where decisions were missed between sessions
- Decision coverage backfill — audited and filled 3 missing decision gaps for recent commits
Unified Architectural Change Detection (v2.7.0)
- Single intelligent watcher — merged two overlapping auto-detection systems into one unified service with better noise filtering and git-diff-based thresholds
- Code file monitoring — now tracks significant changes to services, views, utils, commands, MCP server, core library, and CLI code — not just config files
- No more blocking modals — replaced the old forced decision popup with a non-intrusive 3-mode system (review, immediate, smart)
- Disable mode — new
"disabled" option in autoLogMode setting for full control over change detection
- Duplicate detection — automatically skips files that already have a recent decision logged
- Feature gate awareness — respects Free tier decision limits, gracefully falls back to debt tracking when limits are reached
Previous (v2.6)
Cross-Workspace MCP Config Auto-Migration (v2.6)
- Full workspace coverage — MCP config auto-update now migrates ALL workspace entries (Cline, Roo Code) on extension update, not just the current workspace
- Universal stale path detection — detects and replaces ANY old extension version path in MCP configs, not just the single most-recent previous version
- Zero manual config maintenance — eliminates the root cause of zombie MCP accumulation: stale config files pointing to old extension binaries
Hardened Zombie MCP Prevention (v2.5.8)
- Periodic 15-minute cleanup sweep — background interval catches processes that become zombies mid-session
- Version-mismatch detection (Heuristic 5) — automatically kills MCP servers from old extension versions
- Dormant parent detection (Heuristic 6) — identifies zombie extension hosts using RSS + child age thresholds
Zombie Process Prevention & Reliability (v2.5.7)
- 6-heuristic zombie MCP detection — identifies and kills zombie extension hosts stuck at 100% CPU
- Documentation consolidation — removed ~55 duplicates, organized docs into clean hierarchy
- Code structure refactoring — improved readability and maintainability across core modules
Decision Coverage Auditing (v2.5.6)
audit_decision_coverage tool — compares git commits against logged decisions to find gaps- Startup coverage checks — MCP server scans recent commits on every session start and reminds you about unlogged decisions
- Auto-draft decisions — optionally create draft decisions from uncovered commits with a single flag
- Progressive disclosure memory — context is delivered in layers instead of dumping everything at once, reducing token waste
Background Mode & Git Integration (v2.5)
- Silent operation — Continuity runs in the background with zero interruptions
- Git commit detection — auto-detects commits and reminds you to log related decisions
- Persistent decision reminders across sessions
Multi-Agent Support
- Agent-specific instruction files — AGENTS.md (Codex), CLAUDE.md, .cursorrules, copilot-instructions.md, GEMINI.md
- Tuned per agent — imperative style for Codex/Copilot, suggestive for Claude/Cursor
- OpenAI Codex integration with terminal detection and fallback config
- Full process lifecycle management on Windows (wmic/PowerShell + taskkill)
- Zombie and orphan MCP process cleanup now works on Windows
- Cross-platform signal handling and temp directory resolution
Security & Reliability
- 35 advanced security detection types including multi-turn manipulation, TOCTOU race conditions, and covert channel detection
- 620+ security tests across ~5,500 lines of security code
- Hardened stdio bridge lifecycle for background operation
- SIGPIPE and broken pipe handling to prevent zombie processes
Previous (v2.3)
- Fluent AI Memory — auto-generated instruction files, first-call context injection, pattern-based reminders
- Fixed semantic search cosine similarity scoring
- MCP Tool Profiles — Core (12), Standard (29), or Full (92) tools
Documentation
Support
Memory that belongs to your code — not your model.
| |
