thin-memory(1) · service manual memory for ai agents rev 2026.10 · edge
thin-memory / agents [ claim keyclaim ]

memory for ai agents

Memory that outlives the context window.

thin-memory turns raw text into durable facts, keeps them in isolated containers, and returns a ranked context block whenever your agent asks. It runs on Cloudflare's edge with managed models — one tmk_ key, no vector database, no BYOK.

[ get a free key ] [ read the api ]

25 adds / 500 searches / 10 extractions per month · no card · key in one request

v1 / searchresponse, trimmed
container   user_amanda
plan        free
facts       38 / documents 4

> what does amanda prefer?

  0.91  rust for services, typescript for the ui
  0.88  picks mongodb atlas over postgres
  0.84  leads design at a 12-person studio in lisbon
  0.71  ships on fridays, async reviews only

4 of 38 facts / 41 ms
json trimmed for print · real responses include ids and metadata
01

what it is

a memory layer, not another database

An agent forgets everything the moment the context window closes.

thin-memory is the smallest useful fix: one API that accepts text, extracts the durable facts with managed models, stores them next to their embeddings, and serves them back scoped, ranked and token-aware.

It is deliberately thin. No pipeline to operate, no embedding provider to choose, no vector cluster to size. D1 for facts, Vectorize for the vectors, Workers AI for the embeddings. You bring an agent and a key.

fig. 02 / pipelineone add
   raw text / documents / chat
              |
              v
   +--------------------------+
   |  extract facts           |   managed models
   +--------------------------+
              |
              v
   +--------------------------+
   |  embed + store           |   workers ai / d1 / vectorize
   +--------------------------+
              |
              v
   +--------------------------+
   |  recall + rank           |   container-scoped
   +--------------------------+
              |
              v
   context block  ->  your agent
every call is synchronous · extraction failures never fail an add
02

how it works

three calls, no ceremony

[1] add

Send the raw text.

Documents, chat turns, notes. One call stores the source and asks the extractor for facts.

POST /v1/add
{ "content": "..." }

[2] learn

Keep the facts.

Managed extraction returns short, self-contained statements. Each one is embedded and indexed in your container.

- amanda leads design in lisbon
- prefers rust for services
- chose atlas over postgres

[3] recall

Ask for context.

Search returns ranked facts and documents. Profile returns the stable block an agent should read first.

GET /v1/profile
-> 41-token context block
03

what you don't have to build

the boring parts are the product
04

connect

one URL, one header

Point your client at the MCP endpoint. Keep the key.

The server exposes add_memory, search_memory, context, list_memories and delete_memory. Streamable HTTP, stateless, no session to keep alive.

  • [ claude ]
  • [ cursor ]
  • [ opencode ]
  • [ any mcp client ]

Prefer plain HTTP? The API reference covers the six REST calls and the exact error shapes.

mcp.json 
{
  "mcpServers": {
    "memory": {
      "url": "https://thin-memory.parithosh.workers.dev/mcp",
      "headers": { "Authorization": "Bearer tmk_..." }
    }
  }
}

replace tmk_... with the key from /claim

05

pricing

no seats · no overage invoices
per month free pro
adds2550,000
searches500200,000
managed extractions1010,000
price$0$19 / mo
[ start free ] [ compare plans ]

Your agent will forget this page in ten seconds.

Give it somewhere to put what matters. Free keys are free forever — sign in with a verified email at /claim to mint one. Pro is $19 a month when you need the room.

[ get a free key ] [ buy pro ]
end of page 01 / 06 thin-memory · service manual