🧠 Continuity — Persistent Memory for AI Coding
Stop losing decisions between AI sessions.
AI coding tools are stateless. They forget your architecture, your constraints, and your decisions every time you start a new session.
Continuity fixes that. It gives your AI persistent, repo-owned memory — so it remembers what matters and stays consistent over time.
⚡ Who This Is For
- Developers using Cursor, GitHub Copilot, Claude Code, or Gemini
- Teams tired of re-explaining the same context every session
- Anyone building real systems with AI — not just prompts
❌ Without Continuity vs ✅ With Continuity
| Without Continuity |
With Continuity |
| Re-explain decisions every session |
Decisions automatically recalled |
| AI forgets architecture |
Architecture stays consistent |
| Conflicting implementations |
Conflicts detected and flagged |
| Context resets constantly |
Context persists across sessions |
AI goes from a forgetful intern → a staff engineer with memory.
🎥 See It In Action
Prefer full quality? Watch on YouTube.
Side-by-side: same codebase, same question, identical AI tool. The only difference is whether Continuity is active.
🧩 What It Does
Continuity adds a state layer to your AI tools:
- 🧾 Remembers why decisions were made
- 🔍 Recalls the right context at the right time
- 🛑 Prevents conflicting or unsafe actions
- 🔄 Carries context across sessions and tools
⚙️ How It Works
- Log decisions as you work — manually, via the
continuity CLI, or auto-detected from git commits.
- Stored locally in your repo —
.continuity/decisions.json, committed alongside your code.
- AI tools recover context automatically via instruction files (CLAUDE.md, .cursorrules, AGENTS.md, GEMINI.md) or optional MCP integration.
- Only relevant context is injected — no prompt bloat, no wasted tokens.
🔒 Local-first. No cloud memory. You own everything.
🔑 Core Features
🧾 Decision Memory
Log once — your AI remembers forever.
🧠 Semantic Search
Find decisions by meaning, not just keywords.
🔗 Code Intelligence
Link decisions to files, symbols, and changes. Get refactor warnings when touching linked code.
🔄 Session Handoff
Resume work without losing context. Auto-generated handoff documents capture deltas between sessions.
- Claude Code (
CLAUDE.md)
- Cursor (
.cursorrules)
- GitHub Copilot (
.github/copilot-instructions.md)
- Google Gemini (
GEMINI.md)
- OpenAI Codex, Google Antigravity
- Any instruction-based AI assistant
📊 Knowledge Graph
Visualize relationships between decisions with a force-directed graph.
📥 Source Ingestion
Drop documents into .continuity/sources/ and ingest them into your knowledge base with automatic entity extraction and deduplication.
📈 Reports
Generate summaries, comparisons, timelines, and deep-dive analyses from your decisions.
🧪 Wiki Health Check
Detect stale decisions, contradictions, orphaned records, and coverage gaps.
🔌 MCP Integration (Optional)
Unlock advanced capabilities for MCP-compatible tools:
- Tool interception
- Policy enforcement
- Context injection
- Automation workflows
59 MCP tools across decision logging, search, code intelligence, governance, freshness, reports, and sources.
🚀 Getting Started
- Install from the VS Code Marketplace
- Open your project in VS Code
- Choose your profile — Continuity supports software, writing, research, medical, legal, and general-knowledge profiles
- Log your first decision — your AI will remember it next session
That's it — no cloud setup, no account required.
💰 Pricing
14-Day Free Trial
- Full Pro access
- Unlimited decisions
- Every feature unlocked
- No credit card required
After the trial, your decisions are preserved (read-only). Upgrade to Pro to keep logging.
Pro Plans
| Plan |
Price |
| Monthly |
$9 / month |
| Annual |
$89 / year (save 18%) |
| Lifetime |
$199 one-time |
Upgrade to Pro →
🧠 What This Actually Is
Continuity is not:
- A note-taking tool
- A prompt helper
- Just another extension
Continuity is:
The state management layer for AI coding tools.
Just like React manages UI state, Continuity manages AI reasoning over time.
⚡ Why It's Different
- Local-first — no vendor memory, no cloud dependencies
- Repo-owned — versioned with your code, reviewed in PRs
- Cross-tool — same memory serves every AI assistant you use
- Built for real work — production repos with 1,000+ decisions, not demos
⭐ If This Helps You
If Continuity saves you time, please leave a review on the Marketplace — it helps others find it. 🙏
What's New
v2.20.44 — April 2026
- MCP Config Self-Heal on Upgrade — after every extension upgrade, Continuity now rewrites stale version-pinned paths in
.mcp.json, .vscode/mcp.json, Claude Desktop, Cursor, Copilot, Gemini CLI, Codex, Antigravity, and Claude CLI configs on activation. Previously, upgrading from an older version (e.g. 2.20.6 → 2.20.42) left those configs pointing at a garbage-collected extension directory, and every MCP client silently failed to spawn the server until the user ran a setup command. Now it's automatic.
v2.20.42 — April 2026
- Shorter Name on the Marketplace — the extension card now reads simply "Continuity" instead of "Continuity Ultimate". The publishing ID (
hackerware.continuity-ultimate) and marketplace URL stay the same, so existing installs update seamlessly.
- YouTube Demo Link — the README's "Prefer full quality?" link now opens the side-by-side demo on YouTube (
https://youtu.be/T5xG5e5Wwd0) instead of prompting a raw MP4 download. Plays inline for everyone, no private-repo fallback needed.
v2.20.39 — April 2026
- Side-by-Side Demo on the Marketplace Page — the README now embeds an animated GIF (with a linked MP4 fallback) showing two identical codebases being asked the same question — one with Continuity's memory active, one without. The difference in the AI's response is the pitch.
- Peer-Review Demo Projects —
demo-projects/peer-review/{with-continuity,without-continuity}/ ship as a reproducible reference: identical Express + Postgres codebases (paydash-api), one pre-seeded with 7 architectural decisions, one stripped clean. Anyone can clone, open both side-by-side, and verify Continuity's behavior on their own machine.
- Build-Config Tightening — added
demo-projects/ to the tsconfig.json and webpack.config.js exclude lists so its nested Express/Pino/Stripe imports don't leak into the main extension compilation.
v2.20.37 — April 2026
- New README Positioning — Marketplace page rewritten for clearer value framing: "Stop losing decisions between AI sessions." Before/after comparison, skimmable feature list, sharper differentiation from note-taking and prompt tools.
- NPM Distribution of CLI, Core, and Benchmarks —
@continuity/cli@3.3.4, @continuity/core@1.0.2, and @continuity/benchmarks@0.1.4 are now on the public npm registry. Install globally with npm install -g @continuity/cli for a standalone CLI, or run continuity bench peer-review to execute the cognitive-firewall benchmark on any machine.
- Small-Corpus Search Fix —
SemanticSearchService now uses a lower minScore default (0.05) for corpora under 20 decisions, so synonymy-only matches aren't over-filtered on small projects or fresh installs.
- Benchmark Suite Improvements — Peer-review harness now writes a retrieval-debug block per scenario, applies a fair refusal prompt, and auto-creates the reports directory. The CLI copies every run into
$PWD/.continuity/bench/ so reports live in your project.
v2.20.36 — April 2026
- MCP Scope-Conflict Fix — When a workspace
.mcp.json already defines a continuity server, the setup wizard now skips writing user-scope Claude Code and Claude CLI entries. Eliminates the "conflicting scope" warning Claude Code emits when the same server appears in both project and user scopes.
- CLI
capabilities Command — New continuity capabilities command lists every CLI command, MCP tool surface, and feature flag in one place for quick discovery.
- README Sync — Documentation refreshed to reflect the current CLI surface and recent MCP tool additions.
v2.20.33 — April 2026
- Quieter Session Notes — Removed the disruptive "AI updated session notes" popup notification that fired on every MCP/CLI session update. AI updates now log silently.
- Governance Handler Fixes — Improved conflict detection in governance tools.
- Setup Wizard Improvements — Enhanced SetupWizard with better flow and error handling.
- Token Monitor Fix — Resolved hook failures for Gemini CLI sessions.
v2.20.30 — April 2026
- Test Suite Resolution — Fixed 29 failing tests, resolving regressions in AIInjector, SemanticSearchService, and MCP health monitoring.
- AIInjector Injection Fix — Added missing
injectToCline method to AIInjector, restoring compatibility with Cline-like tools.
- Semantic Search Refinement — Fixed search logic for small corpora in
SemanticSearchService that was broken by aggressive optimization.
- Improved MCP Setup — Hardened MCP setup reliability with better transport mode handling and server path detection.
- Token Budget Monitor — New
token-monitor.sh utility for cross-LLM session handoff, tracking budget and usage across multiple AI sessions.
- UI Responsiveness — Fixed several UI overflow issues in the decision cards and sidebar, ensuring a clean fit at any sidebar width.
v2.20.28 — April 2026
- MCP Stability Fix — Resolved a long-standing issue where the MCP server would disconnect during Claude Code sessions. Removed aggressive shutdown triggers (heartbeat watchdog, parent PID polling, stdin close handlers) that were killing healthy server instances. The MCP server now stays connected reliably.
- Lock Acquisition Fix — Disabled
acquireLock() which was sending SIGTERM to the running MCP server on every Claude Code reconnect, causing chronic disconnections.
- Workspace Detection — Removed hardcoded
WORKSPACE_ROOT from MCP setup configs, allowing dynamic workspace detection.
v2.20.27 — April 2026
- Codebase Search Engine — New unified search blends decision records with project file search (TF-IDF + filename matching). MCP tools:
search_codebase and search_all.
- Answer Synthesizer — Search results now include a natural-language synthesis of matched decisions, improving downstream LLM answer quality.
- Code Context Enrichment — Search results auto-attach relevant code snippets from files referenced in decisions. Opt-in via
includeCodeContext parameter.
- Query Classification — Automatic detection of broad/specific/ambiguous queries with topic clustering for vague searches.
- Compact Wake-up Mode — New
compact: true parameter on get_quick_context returns under 300 tokens for minimal context loading.
- Per-Decision Editor Access — New "Open in Editor" button on each decision card opens
decisions.json scrolled to that decision's exact line.
- Memory Wars Dashboard — Redesigned benchmark comparison page at
benchmarks/memory-wars/index.html with dark editorial design, real MemPalace RAGAS/RGB scores, and 46-2 head-to-head results.
- Real Competitor Benchmarks — Ran RAGAS and RGB against MemPalace directly: faithfulness 0.40 (vs our 0.92), robustness 0.63 (vs our 0.95).
- Precision Overhaul — Bigram-enhanced dedup with lower threshold (0.70), adjusted weights, and backfill to maintain context window size.
- LongMemEval Benchmark — New bench-09 implementing long-term memory recall evaluation (500 queries).
- Real ADR Corpora — Added real-world Architecture Decision Records from open-source projects alongside synthetic corpora.
v2.20.22 — April 2026
- Search Result Deduplication — Context precision improved by detecting and consolidating near-duplicate decisions in search results. Reduces redundant chunks that drag down retrieval quality.
- Small-Corpus Smart Threshold — Corpora with fewer than 50 decisions now bypass the search pipeline entirely, loading all decisions as direct context. Eliminates the overhead penalty on small projects.
- Freshness Grey Zone Fix — Tuned the 5-factor staleness scorer to catch decisions aged 4–8 weeks with
needs-review tags that previously evaded detection.
- Multi-Domain Test Corpora — Added 3 synthetic benchmark corpora (Web API, Data Pipeline, Mobile App) for cross-domain retrieval evaluation, addressing the single-corpus generalizability limitation.
- Multi-Judge Evaluation Framework — RAGAS and RGB benchmarks can now run against multiple LLM judges (Claude Sonnet, Haiku, optional GPT-4o) with inter-judge agreement metrics.
- Benchmark Comparison Page — New
benchmarks/comparison.html marketing page with tabbed dashboard comparing Continuity vs MemPalace across 6 evaluation dimensions.
v2.20.20 — April 2026
- Freshness Scoring Refinements — Improved staleness calculation in the MCP freshness tool with better edge-case handling for decisions without file references or tag history.
- Semantic Search Tuning — Minor precision improvements to the hybrid search pipeline in SemanticSearchService.
- GovernanceLock Fix — Corrected a minor issue in governance lock evaluation.
- Benchmark Data Refresh — Full re-run of retrieval, RAGAS, RGB, and freshness benchmarks against the live 1,675-decision corpus.
v2.20.18 — April 2026
- Decision Freshness Scoring — New staleness detection engine computes a freshness score (0–100) for every decision based on age, file references, supersession chains, tag activity, and status decay. Stale decisions are automatically annotated in
search_decisions and get_quick_context results when they exceed the configurable threshold (default 30).
- Freshness Report Tool — New
get_freshness_report MCP tool surfaces the most stale decisions and suggests Dream consolidation candidates.
- VS Code Setting —
continuity.freshnessThreshold lets users tune the staleness threshold from the Settings UI.
- Instruction File Integration — Generated instruction files (CLAUDE.md, AGENTS.md, etc.) now include a Decision Freshness section with stale-decision guidance.
v2.20.16 — April 2026
- Search Precision Hardening — Reciprocal Rank Fusion (RRF) stabilizes hybrid search by merging semantic and keyword ranks. Score-Gap Adaptive Limiting ("Cliff Detection") auto-prunes low-confidence results. Redundancy deduplication filters semantically duplicate decisions (cosine similarity > 0.80). Lead-sentence specificity boost rewards decisions that directly address the query topic.
- Peer-Review Benchmark Suite — Project Chronos benchmark suite with 5 benchmarks scored by live Claude Haiku 4.5 API (zero mocked scores). Frameworks: BEIR/TREC retrieval, RAGAS generation quality, RGB robustness, token scaling, ANN vector performance. Results: 0.68 context precision, 0.91 faithfulness, 0.95 robustness, 98.1% token savings at 5K decisions.
- Scholarly Evaluation Document — Full academic paper (
benchmarks/ACADEMIC_SUMMARY.md) with 7 references (NeurIPS, EACL, AAAI), comparison against published baselines, threats to validity, and reproducibility instructions.
- Build Hygiene — Benchmarks directory excluded from tsconfig, webpack, and .vscodeignore to prevent build pollution.
v2.20.14 — April 2026
- 7 New CLI Commands —
update, archive, lint, report, source, audit, echo-chamber. All critical features previously MCP-only now have CLI equivalents.
- Settings Tab Cleanup — MCP section replaced with small "Advanced" area. Header MCP dot and label removed. Onboarding buttons no longer push MCP setup.
- Instruction Files Rewritten — Generated CLAUDE.md, AGENTS.md, etc. no longer reference MCP tool calls. CLI commands and grep are primary.
- Pre-commit Hook Improved — Reduced false positive warnings for sidebar case handlers.
v2.20.12 — April 2026
- MCP Demoted to Optional — MCP is no longer treated as required. No more auto-setup on activation, disconnect warnings, zombie cleanup intervals, or setup banners. Instruction files (CLAUDE.md, .cursorrules, etc.) are now the primary context delivery mechanism. MCP remains available for users who want it.
- Removed Cline/Roo Integration — Cline and Roo Code MCP configuration removed. Supported tools: Claude Code, Cursor, GitHub Copilot, Google Gemini, OpenAI Codex, Google Antigravity.
- Clean Status Bar — No more MCP warning icon. Status bar shows domain emoji + decision count only. MCP status in tooltip only.
- MCP Dot Indicator Fixed — Sidebar MCP dot was hardcoded green. Now reflects real connection status (hidden when disconnected, green when connected).
Earlier releases
v2.20.8 — April 2026
- Persistent MCP (HTTP mode) — One-click command (
Continuity: Enable Persistent MCP) starts the MCP server as a daemon on port 3100. Survives VS Code restarts, eliminates zombie processes.
- MCP Settings Panel — Transport mode (auto/stdio/http), HTTP port, enable/disable persistent MCP, daemon status check, and tool profile selector all in the sidebar Settings tab.
- Autolog Hardening — Self-healing decision storage, startup recovery queue, hook durability improvements, CLI fallback behavior fixes.
v2.20.6 — April 2026
- Security & Governance Dictionaries — Comprehensive security terms added to all 5 entity dictionaries (SAML, SSO, TLS, AES, CSRF, XSS, GDPR, HIPAA, etc.).
- Governance Conflict Detection — Security-specific conflict keywords and domain contradiction maps.
v2.20.3 — April 2026
- Zombie MCP Prevention — Hardened heartbeat lock files, faster stdin-close detection, aggressive cleanup of orphaned MCP server processes.
- CLI Improvements —
sync --silent fully suppressed, graceful error when dashboard is missing the blessed dependency.
v2.19.9 — April 2026
- LLM Wiki Pattern — Full persistent, compounding knowledge base with source ingestion, report generation, lint checks, and query write-back.
- Report Generation — 4 formats: summary, comparison table, timeline, deep-dive.
- Wiki Lint — 7 health checks producing a 0-100 health score.
v2.19.4 — April 2026
- Anti-Sycophancy Guardrails — Detects and breaks echo chambers in the decision pipeline.
- Epistemic Rigor Scoring — Evaluates whether decisions document alternatives and trade-offs.
- Contradiction-Aware Search — Search results surface tensions between contradicting decisions.
v2.19.0 — April 2026
- Domain Profiles — 6 domains: software, writing, research, medical, legal, general.
- Engagement Suite — Onboarding cards, memory score, contextual nudges.
- Graph Improvements — Entity details on selection, 12 category system.
v2.17.x — April 2026
- Lazy MCP Startup, Entity-Graph Search, Install integrity verification.
v2.16.0 — April 2026
- Unified trial and license across extension and CLI.
continuity generate command for AI instruction files.- Git hooks for auto-logging decisions on commits.
v2.15.0 and earlier
- Initial CLI with 20+ commands, Dream Engine, Knowledge Graph, decision logging, MCP integration, semantic search, sidebar dashboard.
🔒 Privacy
- All data stays local — decisions, sources, and session notes live in
.continuity/ in your workspace.
- No cloud storage — Continuity never sends your decisions to any server.
- Anonymous analytics — opt-out via
continuity.analytics.enabled: false in VS Code settings.
- API keys stored securely — uses VS Code SecretStorage (encrypted, never plain text).
🆘 Support
| |
