AgenticMind

The auditable, self-improving knowledge & memory layer for AI agents.

Grounded answers with provable citations, a full why-trace for every answer, and a corpus that improves itself — served to any agent over MCP. Zero-key, multilingual, and self-hostable on Postgres alone.

Quickstart · See it work · Agent tools · How it works · Why · The Standard ↗

_{If this is useful, a ⭐ helps others find it — and tells us to keep going.}

Not "memory storage for an agent." AgenticMind is the substrate an agent points at when it needs answers it can trust, a trail it can audit, and a knowledge base that compounds.

Most agent memory is a vector store with save() and search(). That buys you fuzzy recall and zero accountability: you can't tell why an answer came back, whether it's current, or whether a source even supports it. AgenticMind treats knowledge as a first-class, auditable, self-improving substrate — and exposes it to any agent over the Model Context Protocol.

✨ Why AgenticMind

• 📌 Citation-enforced — every claim in an answer is keyed to a numbered source. No source, no claim.
• 🔍 Fully auditable — a replayable why-trace for every answer: what was retrieved, ranked, and used.
• ♻️ Self-improving — validated answers are promoted back into the corpus by a judge-gated compounding loop, driven by programmatic signals (not human thumbs).
• 🧩 Tiered retrieval — chunks → typed fact cards → knowledge graph; hybrid vector + full-text, recency-aware.
• 🔐 Safe by construction — scoped, least-privilege MCP tokens, fail-closed auth, guardrails on input and output.
• 🐘 One datastore — Postgres + pgvector carries vectors, full-text, the graph (recursive CTE), and the durable queue. No Redis, no Neo4j, no vector-DB sprawl.

🔧 How it works

flowchart TD
  A["🤖 Agent"] -->|"MCP request"| R["Tiered retrieval<br/>pgvector + full-text + graph"]
  R --> Y["Citation-enforced synthesis"]
  Y -->|"grounded answer + [citations]"| A
  Y --> T[("Replayable why-trace")]
  Y -->|"programmatic signals"| L["Judge-gated compounding loop"]
  L -->|"promotes validated knowledge"| R

A request comes in over MCP → the engine retrieves across three tiers → synthesises an answer where every claim cites a source → logs a replayable trace → and feeds programmatic signals into a loop that promotes validated knowledge back into the corpus.

🎬 See it work

A real kl_ask_global call against a corpus seeded with the Agentic Product Standard. The question deliberately has two halves — one the corpus can answer, one it can't:

// → kl_ask_global
{ "question": "When should I use a multi-agent architecture instead of a single agent,
                and what must every agent ship with according to the standard?" }

// ← response (trimmed)
{
  "answer": "The provided sources do not specify when to use a multi-agent architecture
             versus a single agent. … According to the Agentic Product Standard, every
             agent must ship with a written Agent Contract [1]. This contract must cover
             ownership, forbidden actions, acceptance criteria, failure modes, escalation
             rules, and logging requirements [1].",
  "citations": [
    { "number": 1, "title": "Agent Contract requirement",
      "materialId": "ba44971b-…", "score": 0.46, "origin": "chunk" }
  ],
  "model": "google/gemini-3.1-flash-lite-preview",
  "retrievalMs": 606, "generationMs": 890,
  "phases": [ {"phase":"embed","ms":552}, {"phase":"retrieve","ms":37},
              {"phase":"synth","ms":890}, {"phase":"output_filter","ms":2} ],
  "telemetryId": "cc942e54-…"
}

Look at what didn't happen. The half the corpus couldn't support, the model refused to answer — "the provided sources do not specify…" — instead of fabricating it. The half it could support is keyed to a numbered citation you can open. And every answer comes with a why-trace (phases, model, telemetryId) you can replay. That's the whole pitch in one call: no source, no claim — and a receipt for every answer.

🆚 How it's different

✅ Use it when / 🚫 reach for something else when

Use AgenticMind when:

• Your agent must answer from trusted sources, and every claim needs a citation.
• You need a replayable why-trace and a single status (supported / partial / unsupported / conflicted / needs_review) you can gate an agent on.
• Disagreeing or stale sources must be surfaced, not silently resolved.
• You want governed self-improvement — not silent autonomous memory mutation.
• You need self-hosting (Postgres-only) and MCP-native access (Claude Code, Cursor, LangGraph, OpenAI/Claude Agent SDK, custom agents).

Reach for something else when:

• You only need simple personalised chat memory (use a memory SDK).
• You want a hosted API / no-code UI today — AgenticMind is self-hosted infra.
• You need SSO / SOC2 out of the box (see the security model for what exists).
• You're optimising for the fastest prototype, not accountable production.

🛠 Agent surface (MCP)

A headless service (apps/server) exposes the engine as MCP tools over streamable HTTP, with fail-closed per-token bearer auth (scoped, least-privilege):

See What counts as knowledge for the Knowledge Unit contract (what may become stored knowledge), Evals & limits for what we measure and what we don't claim, docs/knobs.md for the optional answer-quality knobs (Tier-B faithfulness, contested-sources, answer policy, source trust), and the security model (fail-closed auth, tenant RLS, lethal-trifecta analysis, supply chain).

There is no frontend — the only consumers are agents over MCP. The tool logic is framework-agnostic in packages/shared/src/lib/knowledge/mcp-tools.ts; the host is a ~60-line Web-standard fetch handler served by Node or Bun.

🧱 Layout

packages/shared/src/lib/knowledge/        ← the tiered engine (the product)
packages/shared/src/lib/ai/               ← chat + embeddings (provider-agnostic; local embeddings by default)
packages/shared/src/database/             ← Drizzle schema + queries (Postgres + pgvector)
apps/server/src/{index,mcp}.ts            ← headless MCP host, Node or Bun (agent surface)
apps/worker/src/jobs/knowledge-feedback/  ← Postgres-scheduled compounding sweep

Architecture notes. Agent-first and Postgres-only: the graph lives behind a GraphStore interface (recursive-CTE traversal on Postgres, no extra service), compounding is driven by programmatic signals, MCP tokens are scoped least-privilege, the agent principal is slim, and the host is a headless Node/Bun HTTP server. Retrieval is multilingual by default — local bge-m3 embeddings cover many languages with zero keys; full-text search uses the language-agnostic simple config (configurable per deployment).

🌐 Ecosystem

AgenticMind is the flagship reference implementation of the Agentic Product Standard — the open standard (plus Claude Code skills) for building production-grade agentic products.

See the standard's AgenticMind case study for a layer-by-layer map of how this repo implements the canon.

🤝 Contributing & license

Contributions welcome — see CONTRIBUTING.md. Licensed under Apache-2.0.

🚀 Quickstart

Run it — no clone (~1 min)

Needs Docker (Compose v2.23+) and an OpenAI-compatible key. One command pulls the published images, generates secrets, brings up Postgres + server + worker, and prints a ready-to-paste MCP config — no repo clone, no token minting:

OPENAI_API_KEY=sk-... sh -c "$(curl -fsSL https://raw.githubusercontent.com/Moai-Team-LLC/AgenticMind/main/quickstart.sh)"

The MCP endpoint comes up at http://localhost:3000/mcp, authenticated with a single static bearer (MCP_API_KEY, auto-generated). Point Claude Code / Cursor at it with the Authorization: Bearer <MCP_API_KEY> header.

Embeddings run locally by default — a zero-key, offline, multilingual model (bge-m3) downloads on first use, so retrieval needs no cloud key. Only the synthesis step needs a chat model: OPENAI_API_KEY for OpenAI (the default), or point CHAT_BASE_URL at any OpenAI-compatible endpoint — a local Ollama or vLLM.

Prefer to read before you run? Same thing, explicit (just the deploy/ drop-in, no full clone):

mkdir agenticmind && cd agenticmind
curl -fsSL https://raw.githubusercontent.com/Moai-Team-LLC/AgenticMind/main/deploy/docker-compose.yml -o docker-compose.yml
curl -fsSL https://raw.githubusercontent.com/Moai-Team-LLC/AgenticMind/main/deploy/.env.example       -o .env.example
curl -fsSL https://raw.githubusercontent.com/Moai-Team-LLC/AgenticMind/main/deploy/gen-secrets.sh     -o gen-secrets.sh && chmod +x gen-secrets.sh
./gen-secrets.sh                   # writes DB password + MCP_API_KEY into .env
# set OPENAI_API_KEY in .env, then:
docker compose up -d

From source (development & contributing)

Requires Docker and Node ≥22.18 (or Bun ≥1.3) — the server and worker run on plain Node or Bun.

git clone https://github.com/Moai-Team-LLC/AgenticMind.git
cd AgenticMind
cp .env.example .env.local         # set AUTH_SECRET (+ a chat key OR local Ollama)
./setup.sh                         # picks npm or bun, starts Postgres, runs migrations
npm run dev                        # headless MCP server on :3000  (or: bun run dev)

Verify the build with npm run check (typecheck + tests) — bun run check works too.

In a from-source dev setup the /mcp route is fail-closed and accepts a bearer typ="mcp" HS256 JWT (rather than the static deploy key). The headless server ships no admin UI — mint one with the issuance script (it reads DATABASE_URL + AUTH_SECRET from your .env.local):

npm run issue-token -- --label "claude-code" --ttl-days 365   # or: bun run issue-token --label …
# prints the bearer on the last line — capture it, it is not stored in plaintext

Then point an MCP client at http://localhost:3000/mcp with that token as the Authorization: Bearer … header. (Lint additionally requires Node ≥22.18 — see .nvmrc.)

Note. The local Docker Postgres has no TLS, so .env.example ships DATABASE_SSL=false and DATABASE_URL on host port 5435. For managed Postgres (Supabase, RDS, …) that requires SSL, set DATABASE_SSL=true.