Shared Memory for AI Agents

make
Claude Code Cursor ChatGPT Windsurf Copilot
remember
everything

Harbour gives your AI agents shared memory. Every insight captured, structured, and reused — so your organization's intelligence compounds.

Install in Claude Code
$ claude mcp add harbour --transport http https://mcp.harbour.dev -s user
Harbour Console
// Layer 1 — Real-time GUI
// Phoenix LiveView + Three.js graph

activity_stream: "live events from all agents"
memory_map: "3D force-directed knowledge graph"
session_debugger: "inspect any agent's memory chain"
What you see
Live dashboard
Watch agents fan out across subtasks. See memory files created in real-time. Trace pointer links in the 3D knowledge graph. Filter by session, type, or reuse heat.
Session pipeline
// Layer 2 — Harbour Kernel
// The 6-stage memory engine

:decomposing → :retrieving → :hydrating
   → :running → :synthesizing → :completed
Memory architecture
MemoryFile → MemoryBlock → AddressRecord
// Content-addressed (SHA-256)
// Chunked: 500-1500 tokens per block

PointerLink: derived_from | supersedes | supports
              contradicts | references | related_to
Retrieval scoring
Scoring formula
0.4 × semantic + 0.15 × recency + 0.2 × importance + 0.15 × reuse + 0.1 × pointer_boost
Provider Bridge
// Layer 3 — Adapter-based LLM access
// One interface, any provider

Bridge.complete(prompt, model: "fast")
Bridge.complete(prompt, model: "strong")
Bridge.embed(content)
Model tiering
fast: "decomposition, summarization"
standard: "agent subtasks"
strong: "synthesis, complex reasoning"
Supported adapters
8 providers
Anthropic · OpenAI · Claude CLI · Codex CLI · Copilot · OpenRouter · T3Code · Custom adapters via Behaviour contract
How It Works

three steps to compounding intelligence

Add Harbour once. Every session after gets smarter.

01

connect via MCP

Add Harbour to Claude Code, Cursor, or any MCP-compatible client with a single command. No SDK, no code changes required.

02

agents store memory

As your agents work, they capture insights, analysis, and decisions as structured memory blocks. Each block is content-addressed and linked in a knowledge graph.

03

future sessions retrieve it

When a new session starts, Harbour surfaces the most relevant prior memory and injects it into context. Your agents start smarter, not from scratch.

Why Harbour

intelligence that compounds

Every session makes the next one better. Knowledge never gets lost.

0
wasted tokens
Never re-derive an insight your team already has. Prior analysis is pre-loaded into context.
∞
compounding sessions
Every session stores memory. Every future session retrieves it. Intelligence grows with every use.
N→N
cross-agent intelligence
What Claude Code discovers, Cursor can use. What ChatGPT analyzes, Copilot remembers. Break down agent silos.
Without Harbour
  • ✕ Every session starts with a blank slate
  • ✕ Prior analysis is lost after each conversation
  • ✕ Same questions re-computed across agents
  • ✕ Context windows wasted on repetition
  • ✕ Organizational knowledge stays fragmented
With Harbour
  • ✓ Sessions start with relevant prior knowledge
  • ✓ Key insights pre-loaded into context automatically
  • ✓ Token budgets used for new reasoning, not rehashing
  • ✓ Cross-agent memory sharing out of the box
  • ✓ Intelligence compounds with every session
Integrations

works with your AI tools

Connect via MCP. One protocol, every agent.

Claude Code
Claude Code
MCP native
Cursor
Cursor
MCP config
ChatGPT
ChatGPT
MCP compatible
Codex
Codex
CLI adapter
Windsurf
Windsurf
MCP config
Copilot
Copilot
MCP native
OpenCode
OpenCode
MCP config
OpenClaw
OpenClaw
MCP compatible