Reusable operating instructions, project-specific workflows, and tool notes that agents can load across sessions.
Agents forget. pumaDB remembers.
Durable memory for agents — without the database work. Connect hosted MCP or a server-side API, store JSON rows, and review what your agents remember.
Keep your answers short, and call me Sam.
Got it, Sam.pumadb.remember("preferences", …)
Can you help me draft an email?
Sure, Sam — keeping it short.← pumadb.query("preferences")
Start
Set up in seconds. No database project required.
Pick hosted MCP for agents or REST for server-side apps, then start writing memory.
[01] hosted MCP
OAuth ready
# Streamable HTTP MCP endpoint
>
https://api.pumadb.ai/mcp
# OAuth discovery and dynamic client registration supported
# Add pumaDB as a hosted MCP server
$
codex mcp add pumadb --url https://api.pumadb.ai/mcp
# Create a custom MCP app
>
Settings / Apps & Connectors / Advanced settings
>
Server URL: https://api.pumadb.ai/mcp
# Add a custom connector
>
Customize / Connectors / Add custom connector
>
Leave OAuth client fields blank
>
https://api.pumadb.ai/mcp
Install guide. Connect from Codex, ChatGPT, Claude, OpenClaw, or any client that supports Streamable HTTP MCP.
[02] server-side API
For small server-side apps, keep a named puma_live_* key in a backend or serverless environment and call /v1 directly.
https://api.pumadb.ai
API quickstart and REST reference. Do not put API keys in React bundles, static sites, mobile apps, browser code, public repos, or other client-side surfaces.
Store
A lightweight memory schema for agents.
Use pumaDB for almost any small, durable JSON record your app or agent needs to remember: facts, state, notes, settings, scratchpads, handoff context, or records that make future tool calls smarter. The consolidated remember MCP tool stores common memory types with inert safety metadata.
Repository facts, architecture notes, branch rules, naming patterns, and decisions that should not be rediscovered.
Communication style, formatting defaults, review expectations, and other durable preferences keyed by person or workspace.
Sources, summaries, links, comparison notes, and follow-up questions from investigations that continue over time.
Open threads, blockers, handoff notes, pending commands, and lightweight status records for long-running work.
Resources, code snippets, Markdown, commands, and config examples stored as inert records for review before use.
Safety rails
Reviewable memory with built-in recovery.
pumaDB keeps agent memory deliberately small and explicit: scoped table limits, per-key rate limits, filtered cleanup, natural edits, and automatic version history for updates and deletes.
20 tables, 1,000 rows per table, and 25 MB total storage per account.
30 writes per minute per key and 60 reads per minute per key.
Every update and delete archives the prior row content. The last 10 versions are kept for 30 days and can be restored with one call.
Let agents turn plain-language requests like "remember that I prefer short PR summaries" into filtered updates without creating duplicate rows. Bulk updates require an explicit opt-in.
Small queries return inline JSON. Large text, larger result sets, or includeLink: true create short-lived viewer and download links.
REST and MCP
One memory surface, two ways to call it.
Use REST endpoints from trusted server-side code, or expose the same row operations to agent clients through MCP tool calls.
HTTP routes for backends, Workers, serverless functions, CLIs, and scripts using bearer API keys.
GET /v1/{table}
POST /v1/{table}
POST /v1/{table}/update_row
POST /v1/{table}/update_where
DELETE /v1/{table}
Agent-facing tools for clients connected to the hosted MCP server, authorized through OAuth. Prefer remember for new typed memory writes; the older remember_* helpers remain compatible.