Why Kagura Memory Cloud?
Your AI forgets everything after each conversation. Kagura fixes that — and gets smarter every time you search.
Most AI memory tools are just vector databases with a chat wrapper. Kagura is different — it implements the full LLM Knowledge Base pattern (Karpathy's LLM Wiki) at team scale:
| Approach | Storage | Compounding | Scale |
|---|
| Vector DB / RAG | Embedded chunks | None — retrieve-only | Any |
| Karpathy's LLM Wiki | Markdown files | LLM rewrites pages | Personal (~100 pages) |
| Kagura Memory Cloud | PostgreSQL + Qdrant + Neural graph | Hebbian + Sleep Maintenance | Team / org |
| Feature | Description |
|---|
| Adaptive Memory | Every search automatically strengthens connections between related memories. The more you use it, the better explore() discovers hidden relationships. |
| Hybrid Search | Semantic (OpenAI / self-hosted) + BM25 keyword — 96% top-1 accuracy |
| AI Reranking | Self-hosted (Ollama/vLLM — local, free), Voyage AI, or Cohere — cross-encoder reranking for precision |
| Neural Memory Graph | Hebbian learning builds a knowledge graph in the background. explore() traverses it for serendipitous discovery. |
| Agent Memory Substrate | Beyond a knowledge store: delivery modes (pinned / time-triggered), a server-stamped trust boundary, an agent state lane, and a retrieval-feedback signal — the primitives an autonomous agent loop needs. |
| Agent Control Plane (preview) | Workspace-scoped Agent Registry, subtractive context bindings, agent-bound member keys, lifecycle kill switches, and one-call session bootstrap. Introduced in v0.49.0. |
| 72 MCP Tools | Guide, Memory, Agent Substrate, Agent Control Plane, Neural edges, Contexts, Tags, Files (R2), Analyses (Memory Analysis), Resources, Secrets, Sleep Maintenance, Usage, API-Key Bindings |
| Multi-Provider | OpenAI or self-hosted (Ollama, vLLM — local, private, zero cost) for embeddings |
| Team Ready | Workspaces, RBAC, context isolation, shared memory |
| Web UI | Next.js dashboard — contexts, search settings, member management |
| 5-Minute Setup | ./setup.sh and you're done |
Architecture
Workspace (team/org)
├── Context A ("my-project") ← like a folder
│ ├── Memory 1 ← 3-layer: summary / context / content
│ ├── Memory 2
│ └── Neural edges (Hebbian) ← automatic connections
├── Context B ("learning-notes")
│ └── ...
└── Members (Owner/Admin/Member/Viewer)
LLM Knowledge Base — 5-Layer Implementation
Karpathy's LLM Wiki pattern describes a 5-layer "living knowledge base" — beyond traditional RAG. Kagura implements all 5 layers at team scale:
| Layer | Kagura Implementation | Difference from Karpathy's pattern |
|---|
| Ingest | REST /api/v1/memory, MCP remember, R2 file storage, resource tokens | + binary blobs, + multi-tenant |
| Compile | MCP-as-compile-API — chat agent compiles via structured tool calls (remember(summary, content, type, tags)) + Sleep Maintenance for batch consolidation | Continuous micro-compile (not batch wiki rewrite) — schema-enforced output |
| Index | Triple index: BM25 (keyword) + Qdrant (semantic) + Hebbian graph (relational) — all auto-maintained | No manual index.md upkeep |
| Query | Hybrid Search + AI Reranker + explore graph traversal | Beyond markdown grep — supports semantic + relational queries |
| Enhance | Hebbian learning — every recall() strengthens edges between co-retrieved memories. Sleep Maintenance consolidates periodically. | Background graph evolution (zero LLM cost) vs LLM-driven page rewrites |
Compounding loop: Currently explicit (user/agent calls remember() after synthesizing answers). Auto-write-back of synthesized answers is intentionally opt-in to keep noise low.
Adaptive Memory: Two Search Paths
Kagura separates precision search and discovery into two independent paths, each optimized for its purpose:
recall() ──→ Hybrid Search (semantic + BM25) ──→ [Reranker] ──→ Precise results
│
└──→ Hebbian Learning (background) ──→ Graph edges grow
│
explore() ──→ Graph Traversal (Neural Memory) ←─────────────────┘ Related discoveries
recall() — Precision search. Hybrid (semantic 60% + BM25 40%) with optional AI reranking. Returns the most relevant memories.explore() — Discovery. Traverses the Neural Memory graph to find related memories that keyword search would miss. Opt-in over MCP: ?add_tools=explore or ?profile=full on the URL lists it.- Hebbian learning — Every
recall() silently strengthens edges between co-retrieved memories. No explicit training needed — the graph grows organically as you use the system.
This separation is intentional: mixing graph signals into recall degrades precision (validated via benchmarks). Instead, each path does what it's best at.
Data isolation: All data is filtered by workspace_id → context_id → user_id. Memories never leak across boundaries. Single Qdrant collection with payload filtering.
Tech stack: FastAPI (async) · PostgreSQL · Qdrant · Redis · Next.js 16 · OAuth2 · MCP over Streamable HTTP
Vector backend: Qdrant by default. A single-process self-hosted / CLI / edge deployment can instead run the embedded LanceDB backend — "Kagura Lite" (preview) with no separate Qdrant server (KAGURA_VECTOR_BACKEND=lance, cd backend && uv sync --locked --extra lite). Not for multi-worker / SaaS (LanceDB is single-writer). See Deployment → Embedded Vector Backend.
여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗
