Memory API · MCP · Agent plugins

Memory your agents can cite.

Auld turns conversations, documents and events into dated, sourced memory, scoped to who is asking. Your agent recalls the current facts, what changed, and the evidence behind both — over one REST API or an MCP server.

Get API key Read the quickstart Book a call

Pay as you go from $0 / month
Managed or in your own cloud
REST, MCP and 20 framework recipes

Three calls, end to end

POST /v1/records

curl -X POST "$AULD_API_BASE/v1/records" \
  -H "Authorization: Bearer $AULD_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: crm-4471-2026-09-12" \
  -d '{
    "sourceId": "crm-4471",
    "title": "Renewal call notes",
    "text": "Halden Outdoor wants the renewal moved to 1 November. Budget is now 32k, down from 40k.",
    "occurredAt": "2026-09-12T15:30:00Z",
    "author": "user_ruth",
    "visibility": "account-team"
  }'

# 202 Accepted
{ "jobId": "job_01J9Z6…", "sourceId": "crm-4471", "state": "queued" }

GET /v1/jobs/{jobId}

curl "$AULD_API_BASE/v1/jobs/job_01J9Z6…" \
  -H "Authorization: Bearer $AULD_API_KEY"

# 200 OK
{
  "jobId": "job_01J9Z6…",
  "state": "completed",
  "memories": { "created": 2, "superseded": 1, "unchanged": 0 }
}

POST /v1/recall

curl -X POST "$AULD_API_BASE/v1/recall" \
  -H "Authorization: Bearer $AULD_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "query": "What is the Halden renewal budget?", "viewer": "user_ruth", "maxTokens": 2000 }'

# 200 OK
{
  "memories": [{
    "text": "Halden Outdoor renewal budget is 32k.",
    "state": "current",
    "validFrom": "2026-09-12T15:30:00Z",
    "supersedes": "Renewal budget is 40k (from 2026-06-03)",
    "sources": [{ "sourceId": "crm-4471", "title": "Renewal call notes",
                  "occurredAt": "2026-09-12T15:30:00Z",
                  "excerpt": "Budget is now 32k, down from 40k." }]
  }],
  "tokens": 182
}

Illustrative interface example · set AULD_API_BASE to your endpoint

Same agent, same question

The difference is what it can remember.

A support agent is asked about an account it last discussed three weeks ago. Without memory it guesses or asks again. With a recall step, it answers from the newest source and can show it.

Without memorystateless

User · 2 Oct

“Is Halden still on the 40k budget for renewal?”

Agent

“I don’t have records of previous conversations. Could you share the current budget and renewal date?”

Or, with a pasted transcript dump:

“Yes, the budget is 40k.” — an earlier figure, repeated because it appeared first.

With Auld recallcurrent

User · 2 Oct

“Is Halden still on the 40k budget for renewal?”

Agent

“No — it was reduced to 32k on the 12 September call, and the renewal moved to 1 November.”

currentBudget 32k · Renewal call notes, 12 Sep

supersededBudget 40k · Kickoff email, 3 Jun

Illustrative example · fictional company and people

How the engine thinks

Four ideas behind every recall.

01 · Ingestion

Send what happened, when it happened

Records carry their original time, a stable source id and optional metadata. Sends are asynchronous and idempotent, so retries and re-syncs are safe.

Sending data →

02 · Current state

Newer facts supersede older ones

Extracted memories are reconciled against what is already known. A change produces a new current memory and keeps the old one as dated history.

How it works →

03 · Provenance

Every memory points to its source

Recall returns the excerpt, source id and time behind each memory, so your agent can cite it and your team can trace an answer.

Concepts →

04 · Identity

Recall respects who is asking

Identities and audiences decide which memories a viewer may see. A recall as one customer never returns another customer’s records.

Access control →

Interfaces

A REST API for your backend. An MCP server for your agents.

Call the Memory API from any language with plain HTTP, or connect an MCP-capable client and let the model use memory tools directly.

Records — ingest one or a batch, re-send to correct, delete to forget
Jobs — poll processing state before you depend on new memory
Recall and answer — ranked memories with dated sources, within a token budget
Identities and audiences — scope what each viewer can see
Sources — receive webhook events from other systems

API reference Connect an MCP client

MCP client configuration

{
  "mcpServers": {
    "auld": {
      "url": "https://YOUR-AULD-HOST/mcp",
      "headers": { "Authorization": "Bearer ${AULD_API_KEY}" }
    }
  }
}

Illustrative configuration · host and key are placeholders

MCP tools4

rememberStore a dated note or source for the current user

recallRetrieve current memories with sources for a question

forgetDelete a record and the memories derived from it

memory_statusCheck processing state of recent sends

Integrations

Works where your agents already run.

Each recipe shows where memory belongs in that framework — as a tool, a node, or a fixed step — with example code using plain HTTP. No proprietary SDK is required.

All integrations → · Plugins for Claude Code, Codex, Cursor, JetBrains and Hermes →

Use cases

For agents whose answers depend on history.

Deployment

Run it managed, or keep the data in your cloud.

Start on the managed service. When data residency or procurement require it, an Enterprise agreement can place the database — or the whole engine — in your own account.

Pay as you go · Flex

Managed service

No infrastructure to run. Create a project, get a key, send data.

Fastest path to production
Per-project keys and limits
Usage visible per call

Enterprise

Your database

Memory is stored in a database in your cloud account; the engine runs on managed infrastructure.

Data at rest stays in your account
You control backups and retention
Engine upgrades handled for you

Enterprise

Full deployment

Engine and databases run inside your AWS, Google Cloud or Azure account, operated under agreement.

Network boundary you define
Your identity provider via SSO/SAML
Custom retrieval configuration

Compare deployment models →

Evaluation

Judge memory on updates, time and abstention — not only retrieval.

Public suites such as LoCoMo, LongMemEval and BEAM probe different abilities. Our benchmark pages explain what each measures, how to run it reproducibly, and which baselines make a score meaningful.

Auld has not published its own scores on this site. When it does, the method and raw outputs will be published with them.

Benchmarks overview Methodology

What a memory test should cover

recallFind a fact stated once, many sessions ago

updateReturn the newest value after a change, not the first

timeAnswer “what was true on 1 March?” correctly

abstainSay “not known” when nothing supports an answer

scopeNever answer from another user’s data

Pricing