Memory that outlives the model.
mnemo is a universal, persistent memory layer for AI. Write a memory once from any agent, then recall it from every model, app and tool. Owned by your Solana wallet, not by a model vendor.
- Memories stored
- —
- Vaults open
- —
- Recalls served
- —
- Median recall
- —
Every model is an island.
Your chatbot learns your preferences. Your coding agent learns your repo. Your support bot learns your customers. None of it travels: switch models or vendors and you start again from zero.
- 01Siloed
Each assistant keeps its own memory, locked inside one vendor.
- 02Unlocked
mnemo pulls memory out of the model and into a vault you control.
- 03Shared
One vault, every agent. Write once, recall anywhere, keep it forever.
Write with one agent. Recall with another.
This console calls the real mnemo API running on this server. Teach Atlas something, then ask Bolt, an agent on a different model, and watch the memory cross over.
Demo instance: memories can be wiped on redeploy. Don't store secrets or personal data.
A memory protocol, not a memory feature.
Built-in assistant memory is a feature of one product. mnemo is a layer underneath all of them, with an honest status on every piece.
- 1
remember()LiveWrite from anywhere
Any agent, app or model writes through one REST API. Every memory records which agent wrote it, and when.
- 2
embed()Demo engineIndex by meaning
Memories are embedded for semantic recall. The demo uses a fast hashed lexical-concept index; dense model embeddings are next.
- 3
commit()Roots liveHash and commit
Each memory gets a SHA-256 content hash and each vault a Merkle root that changes on every write or forget. Committing roots to Solana comes next.
- 4
recall()LiveRecall from any agent
Any agent with access ranks the vault against a query, whoever wrote the memory and whatever model it runs on.
Agents on top. Memory in the middle. Ownership on Solana.
Models come and go. The vault stays, and the wallet that signed for it decides who reads and writes. That is the difference between memory you rent and memory you own.
signMessage calls, never transactions.Chat assistants · coding agents · support bots · local LLMs · MCP clients
remember · recall · forget · vault roots · rate limits · holder quotas
Wallet-signed vault ownership · token-gated quotas · root commitments (planned)
Three calls to give any agent a past.
Plain JSON over HTTPS, CORS-enabled, no SDK lock-in. Point any agent at the same vault id and they share one memory.
- POST
/api/v1/memoriesremember - GET
/api/v1/memorieslist + root - POST
/api/v1/recallrank by meaning - POST
/api/v1/askrecall + answer - DELETE
/api/v1/memories/:idforget - POST
/api/v1/vault/claimwallet ownership
import { Mnemo } from "./sdk/mnemo";
const mem = new Mnemo({
baseUrl: "https://your-mnemo-host",
vault: "g_your_private_vault_id",
});
// Atlas, a research agent on model A, writes a memory…
await mem.remember("User is building a Solana trading bot in Rust", {
agent: "atlas",
});
// …and Bolt, a coding copilot on model C, recalls it later.
const { results } = await mem.recall("Which language for the new module?", {
agent: "bolt",
});
console.log(results[0].memory.content);
// "User is building a Solana trading bot in Rust"Memory that belongs to you, not the model.
Portable between vendors, provable on Solana, and deletable for real. The properties a memory layer needs before you trust it with years of context.
Switch models without amnesia
Memories live in the vault, not in a chat history. Move from one model to another, or run three at once: they all read the same past.
Your keys, your context
A vault is bound to a Solana address. A free signature proves ownership. No email, no password, no vendor account.
Vault: w_7xKX…pump
Nonce: 8f02a6d15e9b
Signing does not move funds.Ask in your own words
Recall ranks by meaning and recency, not exact keywords, and tells you which agent wrote each memory.
Every change moves the root
Memories are hashed into a Merkle root. Anyone holding the root can check a memory was in the vault and unaltered.
Delete means delete
forget() removes the memory and recomputes the root. Try it below: no soft-delete, no shadow copy.
Session 1 and session 9,000 share a past
Restarts, new devices, new vendors, new versions: the vault outlives all of them. Context compounds instead of resetting.
- Day 1Session 1 · Model Aremember("prefers concise answers")
- Week 6Session 214 · Model Brecall → same preference
- Year 2Session 9,000 · a model that doesn't exist yetrecall → still there
$MNEMO, the memory token.
The community token behind mnemo, fair-launched on pump.fun. In the live demo API it already does one concrete thing: holding it raises your vault quota.
- Network
- Solana
- Total supply
- 1,000,000,000
- Launch
- Fair launch · pump.fun
- Presale / VC
- None
Published here and on the official X account the moment the coin is created on pump.fun. Until then, any "mnemo" contract you see is not ours.
Tier is computed from your on-chain balance when you sign in with your wallet. Quotas are demo settings and can change.
- 1Get a Solana wallet
Phantom, Solflare or Backpack. Write the seed phrase down offline.
- 2Fund it with SOL
Buy SOL on an exchange and withdraw it to your wallet address.
- 3Open the coin on pump.fun
Use the link on this page only, then connect your wallet.
- 4Check the CA, then swap
The contract address must match the one shown here, character for character.
$MNEMO is a memecoin with no intrinsic value and no expectation of financial return. Only spend what you can afford to lose. Nothing here is financial advice.
From live demo to protocol.
Ticked items exist in this repository today. Everything else is the plan, with no dates promised.
- Phase 0now
Genesis
- Landing + live demo API
- Wallet-signed vault ownership
- Merkle root per vault
- Holder quotas read on-chain
- Fair launch on pump.fun
- Phase 1next
Builders
- SDKs published to npm and PyPI
- API keys + usage dashboard
- Dense embeddings for recall
- Postgres + pgvector storage
- Phase 2later
Anchor
- Vault roots committed to Solana
- Client-side encryption
- Per-agent read/write grants
- Phase 3later
Everywhere
- MCP server for any MCP client
- Agent framework plugins
- Import/export from assistant memories
Straight answers.
Including the ones about what doesn't exist yet.
A memory layer that sits underneath your AI agents and apps, so whatever one of them learns can be recalled by all of them, on any model, for as long as you keep the vault.
Stop starting from zero.
Give every agent you run the same long-term memory, and keep it when the models change.