Give your AI agents
shared memory.
An open protocol built on attunement ~ where agents declare who they are, and the right context finds them automatically.
Compatible with MCP and A2A · Progressive adoption from Level 0 to 3 · Open specification
The Problem
Our agents can call tools.
They can talk to each other.
But they can't remember.
They can't share what they know. And when two agents contradict each other, there's no standard way to catch it.
Every multi-agent system today carries its own ad-hoc memory layer ~ bespoke, fragile, and non-interoperable. Akashik is the protocol that replaces it.
Where it fits
MCP handles agent-to-tool. A2A handles agent-to-agent. Akashik handles what they share, how they stay coordinated, and how they resolve contradictions.
What makes it different
Intent is mandatory.
You cannot write to the Field without declaring why. Every finding carries the question it answers. You get a reasoning chain for free.
Attunement, not search.
Agents declare who they are ~ role, task, budget. The Field figures out what's relevant and delivers it. Context finds the agent.
Conflicts are expected.
When agents contradict each other, Akashik detects it and gives you structured resolution paths ~ from last-write-wins to human escalation.
v0.3 · memory that survives a restart
import { createField } from '@akashikprotocol/core'
import { createFileAdapter } from '@akashikprotocol/core/file'
const adapter = createFileAdapter({ path: './field.jsonl' })
await adapter.init()
const field = createField({ adapter })
// Intent is mandatory — the why, not just the what
await field.write({
entry: { topic: 'market-size', cagr: '23%' },
intent: 'validated against three analyst reports',
agent: 'researcher',
})
// Restart the process. The field is still there.
const chain = await field.replay({ topic: 'market-size' })
console.log(chain[0].intent)
// → 'validated against three analyst reports'
// Same field, new process. Nothing was lost — including why it was written. One field, a file adapter, and a restart. replay() proves nothing was lost.
Progressive adoption
Start with Level 0 in an afternoon.
Grow to production-grade when you need it.
REGISTER, RECORD, ATTUNE. In-memory. No embeddings.
~ an afternoon · v0.2 releasedDEREGISTER, RETRACT, SUPERSEDE, REPLAY. Persistent adapters, Lamport clocks, and conflict detection.
~ a week · v0.3 releasedSemantic attunement, MERGE, push subscriptions, REPLAY.
~ serious buildSecurity model, authority hierarchy, coordination extension.
~ production-grade