Memory · what your agent knows

Your agent's memory, queryable in milliseconds.

Fast Scoped Cited

Small, self-generated facts your agents write as they work — shared across every agent, or locked to one, returned with the source they came from.

Get Started two calls, no infra
Pricing
the loop write → recall
coding-agent writes
stored · shared
FACT conf 0.98 fact_a91c

Deploys must run pnpm build before vercel deploy — the CI image has no npm.

source: github.com/acme/web · ci/deploy.md
recalled next request
review-agent recalls
every agent can read
cited ✓
Two calls, one loop
→memory.write(fact)
←memory.recall(query)
No separate indexing step. No batch job.
Shared or locked
shared locked
Your call — every agent, or just one. Nothing leaks past who you allow.
Cited, not guessed
source: ci/deploy.md
Agents quote where a fact came from, not nowhere.
Use cases

Notes agents leave for each other.

One agent writes something down mid-task; any agent you allow picks it up later — no re-work, no re-running the same search.

Coding agents

Agents that don't repeat mistakes. Store a fix or a convention once; every agent touching that repo recalls it before acting.

fixes & conventions, recalled
Long-running workflows

Multi-step agents write intermediate findings and pick them back up hours — or sessions — later.

resumed across sessions
Multi-agent fleets

One agent's research becomes another's starting point — without re-running the same search.

handed between agents
Support & ops agents

Save resolved-ticket context so the next agent on the same issue isn't starting cold.

resolved-ticket context, kept
Onboarding new agents

Point a new agent at existing memory and it inherits everything already written down.

+inherits day-one memory

One memory,
every agent.

Written once, recalled anywhere you allow — with the source attached.

Why Ninelayer

Memory, not a database to run.

Three things a raw key-value store won't give an agent — shown, not claimed.

Fast

Written this request, recalled the next.

A write-then-recall loop built for the request/response cycle of live agentic workflows — not a batch lookup you wait on.

→
memory.write({ fact, source })
req · 1
stored — nothing to index
←
memory.recall(query)
req · 2
returns inside the same request cycle·no batch job to wait on
Scoped

Share it, or lock it. One switch.

+9
every agent can read
✓scope: shared · read = all agents
Cited

Every fact carries its source.

FACTconf 0.96

Staging DB resets nightly at 02:00 UTC — don't seed before then.

source: runbooks/staging.md

Agents recall where something came from — not just what it was.

How it works

From a written fact to an instant recall.

Three steps, two calls. You choose who can read it; it comes back with its source.

1 Write

An agent notes a fact — and where it came from.

Mid-task, an agent writes down something worth keeping, optionally with a source.

coding-agentwrite
memory.write({
  fact: "use pnpm, not npm",
  source: "ci/deploy.md"
})
→
2 Share or lock

Visible to every agent you allow — or just one.

The fact is stored and scoped by one setting. Nothing stays visible beyond who you allow.

◎ shared all agents
◐ locked one agent
→
3 Recall

Any allowed agent calls it back — source attached.

A different agent recalls the fact on its next request, with the source it came from.

FACT conf 0.98

Use pnpm, not npm, for deploys.

ci/deploy.mdrecalled ✓
Get started

Give your agents a memory.

Two calls. No memory infrastructure to run yourself — write a fact, recall it on the next request.

self-serve · no card required
the whole api
→memory.write(fact)
←memory.recall(query)
runs over MCP · mcp.ninelayer.in