Most AI tools have brilliant assistants with amnesia. Forge has memory.
Free forever · local · no account needed · vector database · knowledge graph · confidence decay · contradiction detection · 25 personas · 190+ skills · 150+ commands
You've been building with an AI assistant for weeks. It knows your stack. It knows your decisions. It knows why you chose Postgres over MySQL, why you killed that microservice, why that edge case exists.
Then one day - it doesn't.
It suggests the pattern you explicitly rejected. It forgets the constraint your CTO flagged. It confidently tells you something that used to be true but isn't anymore.
That's not an AI problem. That's a memory problem.
And until Forge, nobody solved it.
Every other tool stores memories forever and treats a note from 3 months ago the same as something you said this morning. Forge doesn't.
The moment Forge receives a memory that conflicts with an existing one, it flags both. You see the conflict. You resolve it. Your AI works from verified facts - not a pile of stale notes it has to guess from.
Every memory has a confidence score that decays over time. A fact you verified today scores 1.0. The same fact 6 weeks later, unverified, scores 0.6 - and your AI tells you before using it to make a decision.
Forge Anvil is a vector database embedded directly in your project - zero setup, no API keys. Every memory is embedded and searched by meaning, not keywords. Ask "what did we decide about auth?" and it finds the answer even if those exact words never appear.
When Forge retrieves a memory, it gives you the fact plus confidence score, status, verified date, age, and any warnings. Your AI knows not just what to say - it knows how sure to be.
A pure JSON store embedded in your project. No server, no Docker, no API key, no vendor to trust. If forgeguru.ai disappeared tomorrow, your Anvil keeps working. Your machine is the source of truth.
Forge Anvil store: Apache 2.0. TF-IDF embeddings: pure JS, zero deps. Every line of the memory engine is auditable. No black boxes. No "trust us." Read it, fork it, audit it - the source is the truth.
Everything your AI knows - visible and controllable from the terminal. Strike, draw, seek, temper, fractures, assay, cast. No magic. No mystery.
Forge doesn't just store memories - it links them. ADRs, decisions, constraints, and retro learnings form a traversable knowledge graph. Ask "what depends on this decision?" and trace the full causal chain across your project history.
Every Forge project is staffed by a roster of typed personas - each with a pre-flight checklist, an explicit output contract, and a defined scope. Solo developer wears all hats sequentially. Collaborators own hats individually. The roster is canonical; headcount is variable.
Every persona enforces at the tool level — not just a prompt preamble. An @analyst literally cannot call Edit on source files. Pair two roles with --pair for adversarial review.
Blacksmith 5-Tier Ladder — earned, not assigned
Every protocol in Forge operates inside the constraints of three apex protocols. No other protocol can control, coordinate, or override across protocol boundaries — only these three can. Together they form the load-bearing triangle of the system.
All other protocols defer to this triangle for authority, sequencing, and record-keeping.
Root control over everything else. Can override, impersonate any persona, nuke anvils, bypass gates, and force-clear bands. The only protocol that sits above other protocols — nothing in Forge operates outside its reach.
The 8-phase spine every project runs inside. Brief → Survey → Strike → Shape → Temper → Quench → Stamp → Ship. Without it, no other protocol has a sequence to attach to. Blacksmith governs it; lifecycle sequences everything below.
The record that survives even a nuke. ADRs and decision logs (DEC-NNN) outlast any project, persona, or restore operation. You can restore an anvil — you cannot un-make a recorded decision. The institutional memory that makes the system learn.
Install the CLI, kindle your Anvil, smelt existing docs into memories, and strike your first ground truth.
Four commands. Your AI has memory. It never confidently uses information it shouldn't trust.
Apache 2.0 · Node 20+ · macOS / Linux / Windows
| Age | Status | Confidence | AI Behavior |
|---|---|---|---|
| 0-7 days | Fresh | 0.9 - 1.0 | Used directly |
| 8-30 days | Aging | 0.7 - 0.9 | Used with timestamp |
| 31-60 days | Stale | 0.4 - 0.7 | Surfaced with warning |
| 60+ days | Archive | < 0.4 | Flagged, not used |
We invented the half-life of AI memory. No other AI dev tool does this.
One hard gate: verify → ship. Everything else is advisory. Single-gate methods get respected.
80+ B-* audit dimensions collapse into exactly 6 named bands - designed to fit Miller's 7±2 so a tired operator can read the verdict at 6am without translation. Anvil verdict = worst band. Three-color system: GREEN ≥85, YELLOW 70-84, RED <70.
Anvil verdict = worst band score. A GREEN product with a RED WEAR band is a RED anvil.
These are not marketing spin. They are architectural decisions baked in from day one.
No other AI dev methodology has a local, anti-hallucination memory DB baked in. Every other tool - BMAD, Cursor, Copilot Workspace, Devin - treats memory as a scratchpad. Forge treats it as infrastructure. Memory with confidence decay, conflict detection, and a knowledge graph is not something you bolt on later. It has to be designed in from day one.
"Forge remembers your project accurately - it verifies, ages, and conflicts-checks every memory." Every developer has been burned by an AI that confidently used stale context. Forge is the first methodology that solves that structurally, not just by asking the model to "try harder."
Ground truth lives on your machine, not a black-box cloud. Every memory Forge stores is a plain file on your laptop - readable, exportable, and deletable without asking permission. Apache 2.0 means the entire memory engine is auditable. The AI your security-conscious colleagues will actually approve.
Apache 2.0 means you can fork the whole thing, swap the embedding engine, extend the knowledge graph, or build adapters for new AI hosts. The methodology is plain text. The writs are versioned files. Nothing is locked in a proprietary format. You own the stack.
Built on Claude Code. Adapters for Codex CLI, Gemini CLI, GitHub Copilot, Cursor. Per-model overlays for Opus, Sonnet, GPT-5.4, o-series. The methodology stays the same; behavior adapts to your host and your model.
The learning loop closes automatically. Past pitfalls inform new reviews. Every retro produces typed learnings with confidence + decay that future review-phase personas auto-query. Loop closure rate ≥10% tracked. Your methodology becomes your moat.
Destructive operations in Forge require earning, not just having, authority. The Blacksmith subsystem is a sudo-style, TTL'd authority layer with two-key safety, full audit trail, and automatic expiry. Authority never silently falls through. Authority never propagates to sub-shells.
Both environment variables must be set simultaneously to authorize destructive operations. Neither alone is sufficient. This prevents accidental automation from triggering irreversible actions.
Audit log: ~/.forge/blacksmith/audit.log · mode 0600 · 15+ event types · survives anvil deletion · forensic record persists permanently.
Seven named anti-patterns with associated band triggers. If your project is heading toward one, the audit catches it before it becomes a crisis.
No commits, no phase progression in 14+ days. The project has gone cold.
Soft-pass audits. Tests exist but don't actually verify the acceptance criteria.
Shipping features outside the original brief without an ADR to record the decision.
RUNBOOK and README last updated 90+ days ago. Operator working from memory.
5+ active projects, none shipping. Portfolio spreading attention below the threshold of progress.
Shipped without running the verify phase. fship-safe blocks this - unless bypassed.
Same anvil, no forward progress. Effort spent without moving the phase needle.
AI-first tools give you speed but no structure or memory. Spec-driven methods (BMAD, GSD) give you structure but still no memory. Forge is the only methodology that ships all three.
| Capability | Vanilla Claude / Cursor | BMAD / GSD | FORGE |
|---|---|---|---|
| Ground-truth memory | None - resets every session | None - docs drift fast | Anvil: confidence decay + contradiction detection, local JSON |
| Memory health dashboard | None | None | forge anvil assay - count, stale %, conflicts, confidence map |
| Session handoff | Context lost on close | Manual notes | quench / kindle / where - structured exit, cold resume in <60s |
| Hard gates | 0 | 5-10 (slow, bypassed) | 1 - verify then ship |
| Typed personas | None | Roles, not enforceable | 17 canonical + Blacksmith apex persona |
| Cross-model audit | None | None | GPT reviews Claude - 15-20 catches per plan |
| Adaptive gating | None | Always-on or skipped entirely | Zero-findings personas auto-skip, hit-rate tracked |
| Learning loop | None | Retros, not auto-applied | Typed learnings, decaying, auto-injected into future reviews |
| Host portability | Tool-locked | Platform configs, no adapter system | Claude / Codex / Gemini / Copilot / Cursor via adapter transforms |
| Audit dimensions | None | Manual checklists | 80+ B-* dims - 6 bands - one verdict |
| Operational knowledge base | None | Agent role files | 41 writs - versioned, cross-referenced, CLI-accessible |
| Reusable skill library | None | Agents, not portable | 190+ certified skills, multi-host slash-command handlers |
| Rebuild resilience | None | None | GURU: 3-zone spec, Cold Iron drills, frequency ratchet, cron-verified |
| Egress / security hooks | None | None | Allowlist + deny-unreviewed-skills + HITL checkpoint |
| Compliance evidence | None | Manual | SOC 2 pack, EU AI Act, SLSA Build L2 |
| Root authority model | None | None | Blacksmith: two-key safety, TTL, full audit log |
| CLI depth | None | None | 190+ commands across 3 tiers |
| Open source | Closed / proprietary | MIT / open | Apache 2.0 - read it, fork it, trust it |
Forge ships with a structured library of writs — versioned, cross-referenced operating protocols covering every function a modern AI-native team needs. Not templates. Not starters. Each writ ships with an Implementation Kit, 30-Day Adoption Playbook, Maturity Scorecard (L0–L5), Worked Example, and CI-runnable Audit Fixtures. Every protocol is v1.0.1 and points to the Forge Pyramid for authority, sequencing, and record-keeping.
| Category | What it covers | Protocols |
|---|---|---|
| forge_writ Pyramid | The 3 apex protocols that govern all others — authority, structure, permanence | blacksmith · lifecycle · decision |
| operation_writ | Forge operational layer — personas, agents, skills, knowledge, retro, nuke | forge · agent · brain · skills · knowledge · operator_self · retro · nuke · cli_reference |
| technical_writ | Engineering — ops, testing, AI, data, architecture, release, migration, backup | 13 protocols |
| platform_writ | Mobile and design — iOS, Android, product design | 3 protocols |
| compliance_writ | Legal and regulatory — security, audit, legal, compliance | 4 protocols |
| marketing_writ | Growth and brand — marketing, brand, communication, research, sales, customer | 6 protocols |
| business_writ | Finance and operations — finance, finops, support, crisis, people, startup | 6 protocols |
Forge ships with a complete security and compliance stack - not bolted on, baked in. 150+ dedicated skills for regulatory automation, supply chain integrity, and enterprise readiness.
All memory lives in your project directory as a plain JSON file. No server, no vendor, no exfiltration risk. We architecturally cannot see your project's context.
Forge generates real compliance artifacts, not documentation theater. The audit log is forensic-grade. The evidence pack is dated and integrity-verified.
Every release ships with cryptographic provenance. NPM packages are signed. SBOM is generated and attested. Supply chain attacks are a known class - Forge treats them that way.
Two binaries. Seventy-eight aliases. Three access tiers. Tab completion across every command, subcommand, project slug, story ID, and persona name.
Forge ships today with Claude Code, Codex CLI, and Gemini CLI. The roadmap extends it into every corner of the developer workflow - IDE, project management, deployment, mobile, and community.
Claude Code + Codex CLI + Gemini CLI. 150+ commands. 17 personas. Forge Anvil memory. 41 writs. 6-band audit. GitHub Action. Astro docs site. CI/CD pipeline.
Linear ticket cross-linking. Notion connector. VS Code extension with NICHE/ROI sidebar. Cursor MDC rules (shipping now in preview). GitHub Copilot adapter.
Vercel + Supabase + Fly.io one-command provision. Obsidian plugin. Slack and Discord apps. Raycast extension. Smithy community platform and certification ladder (Smith $399 → Master Forger $14,999).
Public REST API. Webhooks bridge. Plugin marketplace with 3-tier quality system. iOS and Android native apps. forgedev.ai developer community hub.
If a memory is stale, you'll know. If two memories conflict, you'll see it. If a fact hasn't been verified in 60 days, it won't silently poison your next decision.
Other tools give you a smarter AI.
Forge gives you an AI you can trust.
Forge is Apache 2.0 - read it, fork it, trust it. Install in 30 seconds. Your memory lives on your machine. No account, no subscription, no black box.
Want updates on new releases and community news?
No spam. Unsubscribe anytime. Apache 2.0 - read the source.