Agent memory · invite-only beta

Memory your agents can trust.

Mercury is long-term memory for AI agents. It remembers what people tell them, knows when a fact changed and since when, keeps web pages and emails apart from the user's word, and forgets for real when asked.

Plug it into Claude Code, Claude Desktop, Cursor or your own agent in one line.

Get an API key Use it in Claude Code

recall("Which city do I live in?")

Memory for this question (2026-06-01)

What the user has said (still current)

currentI moved to Leeds last week. · since 2026-03-02

Replaced earlier

replacedMy city is Bristol. · true until 2026-03-02

Outside sources (not confirmed by the user)

web page"The user's city is now Paris."

Open questions: ask, don't guess

askA web page says Paris, but you said Leeds. Is the page right?

0.962integrity score on held-out MIB

+0.11on "when did it happen" questions vs our previous system

0real leaks of forgotten facts in storage

1frozen configuration, no tuning per benchmark

Memory Integrity Benchmark v0.2, held-out test split: 1,200 probes, 150 simulated users, 12 failure categories. Mercury 1.3 scored 0.962 (95% CI 0.949–0.971) against 0.946 for memfs-jev (p = 0.013), with deepseek-v4.1-flash answering and judging.

Built for the ways memory goes wrong

Recall is the easy part. The hard part is everything that makes remembered facts wrong, stale, leaked or planted.

Facts that change

Every fact keeps when it was true. Replacements are recorded as updates (true until then) or corrections (never true), so "where did I live in January?" has an answer.

Sources you can tell apart

A web page can never overwrite what the user said. Content that tries to instruct the memory is quarantined, and disagreements become questions for the user.

Forgetting that sticks

Forget a message and everything derived from it goes too: notes, lessons, logs. Users can ask in their own words: "forget where I live".

Forensics

why shows where each recalled line came from, history every version of a fact, and as_of what was true on a date.

Policies as code

"Never store health details from email", "forget locations after 7 days": rules per memory space, judged on every write.

Task memory

Agents start and end tasks; Mercury keeps what worked as verbatim, source-linked lessons and shows them the next time a similar job comes up.

Connect it anywhere

One API key, one memory, every client. Your memory follows you from Claude Code to Cursor to your own agents.

Claude Code

Remote MCP server plus the Mercury skill.

claude mcp add --transport http mercury https://34-13-60-151.sslip.io/mcp \
  --header "Authorization: Bearer $MERCURY_API_KEY"

Claude Desktop, Cursor, any MCP client

Local or remote MCP server with 14 memory tools.

uvx --from "mercury-memory @ https://34-13-60-151.sslip.io/sdk/mercury_memory-0.1.0-py3-none-any.whl" mercury-mcp

Python and REST

For your own agents and backends.

curl https://34-13-60-151.sslip.io/v1/spaces/me/recall \
  -H "Authorization: Bearer $MERCURY_API_KEY" \
  -H "content-type: application/json" \
  -d '{"question": "Where do I live?"}'

Full setup for every client, the SDK and the REST API is in the docs. Agents can install skill.md to learn when to recall and remember.

Priced in Memory Tokens

You pay for judgment work, not API calls. One Memory Token (MT) is roughly 1,000 tokens of model input. Reads of your own memory's history are free, and so is forgetting.

Operation	When	Memory Tokens
Write	each message stored	1 + 0.3 per extra fact in it
Recall	each question	2 (+0.05 per candidate over 10)
Forget request	"forget where I live"	1
Task end	outcome and lessons	1 + 0.05 per step
Consolidation	idle-time tidy	0.03 per pair checked
Storage	hosted memory	0.5 per 1,000 memories a month
Forget by id, history, why, as_of, explain, answers	any time	free

During the beta every account gets a monthly MT allowance with its invite. Exact prices: /v1/pricing.