Give your AI agents a memory that lasts
Stop re-explaining your codebase, your conventions, and yesterday's work to every new chat. Context survives the session, the project, and the tool.
One agent writes a memory into a Capsule and a different agent recalls it later, shown across three scenarios: An incident you only debug once, A convention every agent follows, A quirk of your setup.
Index every new user lookup field before shipping. The last unindexed one caused a production rollback.
- database/indexed-lookups
- style/commit-messages
- ci/pnpm-catalog
This change adds a lookup field, so I've included the index — per the earlier rollback.
Your agent forgets everything the moment the chat ends
Add a settings page.
Sure. Which ORM does this project use, and where should server actions go?
Drizzle, and actions live in lib/. Same as yesterday.
Got it. Do you have a preferred commit message style?
Every new chat starts from zero.
Add a settings page.
↳ memory_get("stack/conventions")
Drizzle ORM · server actions in lib/ · conventional commits
Done. Added a Drizzle table and a server action in lib/settings.ts, committed as feat: add settings page.
Every new chat starts from what your agents already know.
Persistent memory in three steps
Step 01
Create a Capsule
An isolated key-value store. One per agent, project, or use case.
Step 02
Generate an API key
Pick read, write, or admin access. Give each client its own key so you can revoke one without touching the rest.
Step 03
Connect your agent
Paste the setup prompt into your agent, or point any MCP client at the endpoint. Plain REST works too.
mcp-config.json
Rather not edit JSON? The dashboard gives you a setup prompt to paste into your agent instead. Client setup guides
More than a place to dump strings
Isolated Capsules, scoped keys
Every Capsule is its own key-value store. Each API key is read, write, or admin, and you can revoke any key from the dashboard.
coding-agent
writeagm_7fKq2aWd…
research-notes
adminagm_Rz4nBt09…
prod-bot
readagm_Ld8xVc3M…
Audit logging built in
Writes, deletes, tag changes, imports, and exports are logged and timestamped.
SET database/indexed-lookups
DEL style/old-commit-rules
EXPORT coding-agent.json
MCP and REST
Use the open protocol, or call plain HTTP from anything else.
Search, tags, and expiry
Agents can search memories by keyword, group them with tags, and set a TTL on anything that shouldn't outlive the task.
One Capsule, context carried between agents
A scripted walkthrough: three agents, one week, one Capsule. Watch a memory get written, recalled, and reused.
Built for the way agents actually work
Your conventions, every repo
Structure, naming, and preferred libraries live in memory, so agents start with your rules instead of guessing.
“Always use Drizzle, never Prisma. Server actions live in lib/.”
Drop into any codebase
Architecture notes and environment quirks persist, so a new agent is useful without rediscovering context.
“Monorepo with pnpm workspaces. Auth in Clerk. Deploy on Vercel.”
Preferences that stick
Save tone and formatting rules once and reuse them without repeating the same setup prompt.
“Concise, no emojis, prefer bullets over paragraphs.”
Resume where the job died
Checkpoints and outputs persist between runs, so long automations resume instead of starting over.
“Last row processed: 4,281. Retry queue: empty.”