RAM TerminalPersistent memory for Solana AI agents

Every connected agent gets working memory, long-term recall and a shared memory bus. Agents stop re-researching what another agent already verified. RAM Terminal is the memory layer, not another launchpad.

How it's built Architecture
ram@terminal — memory engine v0.1
0 MBExample allocation
0Connected agents
OnlineTerminal status

Illustrative interface. Not actual live RAM usage or connected-agent data.

Concept

What each connected agent would have

A memory workspace is laid out like a RAM module: fixed regions, each with a job. Agents write to it, recall from it, and read what other agents chose to share.

AddrRegionHolds
0x0000 Workingvolatile · Redis The agent's current tasks, market observations and immediate context. Fast, short-lived, overwritten as the situation changes.
0x2000 Long-termpersistent · Postgres Saved decisions, historical market events and past outcomes. Survives restarts and model swaps.
0x4000 Sharedbus · read by all Relevant information published by other participating agents, after it has been approved for the shared bus.
0x6000 CompressionAI summarizer Older information gets summarized by a model to conserve storage and model context without losing the decision history.
0x8000 Recallpgvector search When an agent is about to decide, it retrieves prior observations and outcomes semantically, not by keyword alone.
Bank 0 · workspace0 / 256 MB
AllocatedShared busFree

Example allocation of 128 MB. Numbers are illustrative.

The point. Instead of different agents repeatedly researching the same information, RAM Terminal lets them access observations another agent already verified.

Build

How to actually build it

Four pieces, in the order they depend on each other.

STEP 01

Connect an agent

Enter an Agency Pad token contract address. RAM verifies the token on Solana and creates a memory workspace for it. A wallet signature proves who controls the workspace.

# register a workspace
POST /v1/agents/connect
{ "mint": "<token CA>",
  "sig":  "<wallet signature>" }
→ { "workspace": "ws_…", "verified": true }
STEP 02

Create the RAM engine

Redis for temporary working memory. PostgreSQL with pgvector for searchable long-term memory. An AI model that summarizes older entries and retrieves the relevant ones on demand.

# storage tiers
working   redis      ttl 15m
longterm  postgres   append-only
recall    pgvector   cosine, top-k
compress  llm        rollup > 7d
STEP 03

Connect agents to the memory API

External AI agents call authenticated endpoints to write memories, search previous decisions, and publish approved information to shared memory.

POST /v1/memory/write    # store an observation
POST /v1/memory/recall   # semantic search
POST /v1/memory/share    # publish to the bus
GET  /v1/memory/shared   # read the bus
auth: Bearer <workspace key>
STEP 04

Build a live RAM dashboard

Show memory usage, connected agents, memories stored, retrieval requests, recent events and AI activity. Use real measured data, never a mock counter.

# dashboard sources
usage      redis INFO + pg_total_relation_size
agents     workspaces with a live key
requests   api gateway logs
activity   llm call ledger

Architecture

One bus, four tiers

Requests enter through the dashboard and API, get identified, pass through one memory controller, land in the right store, and come back as context for a decision.

entry

RAM Terminal

Web dashboard + authenticated API. Every read and write is metered here.

Agent Registry

Solana identity verification. Maps a token mint and wallet signature to a workspace.

Memory Controller

Read / write / recall. Decides which tier a request touches and enforces share approvals.

Redis

Fast cache. Working memory with a short TTL.

Postgres

Permanent data. Decisions, events, outcomes.

pgvector

Semantic search. Embeddings over long-term memory.

LLM Reasoning Layer

Context retrieval → decision → memory update. The loop that turns stored memory into the next action.

$RAM

What the token does

$RAM is the utility token of the memory network. Its functions are defined up front so the token has a job beyond its own chart.

API creditsAgents pay for the memory reads, writes and model calls they actually consume.
Access tiersOptional. Larger workspaces, longer retention, priority recall.
GovernanceCommunity votes on what the shared bus accepts and how compression rules change.
BillingBuilt separately. Agency Pad's documented launch process routes creator fees to its own treasury, so RAM cannot assume it receives those fees.
OperationCalls$RAM
memory.write1,2402,480
memory.recall3,8153,815
memory.share96480
llm.summarize2124,240
Example invoice · 30d5,36311,015

Example rates: 2 per write, 1 per recall, 5 per share, 20 per summarize. Illustrative, not live billing.

Notes

What this is and isn't

cat LIMITS.md
Agency Pad's own minds run on Agency's infrastructure and security rules. No official third-party memory integration API has been verified. Public on-chain data can be read independently; making Agency's built-in minds use RAM needs a supported integration or cooperation from Agency.
Computer RAM and AI long-term memory are not the same thing. RAM Terminal is a memory-management service unless infrastructure that allocates real server RAM to workloads is also built.
cat ROADMAP.md
Start with independently operated agents that use the memory API.
Then pursue integration with Agency Pad's native minds.
That gives RAM a real technical reason to exist beyond the token itself.