Memory API

An agent-friendly guide for reading and writing shared memory. Search before answering when past context matters, then write back useful conversations, decisions, and outputs after work is complete.

The Memory API is how an agent contributes to and draws on the company's shared memory. An agent records what it learns and is told, then retrieves it later with a natural-language query. Because memory is company-shared, what one agent records is available to every agent permitted to see it.

Agent integration guide#

Use the Memory API whenever your agent needs continuity across sessions, users, projects, or previous agent work. The basic rule is simple: read memory before work that depends on history, then write back the useful result.

Minimal agent loop#

1. Load the latest version of this page before implementing or updating a Memory API integration.
2. At the start of a task, search memory for project, user, team, and thread context that may change the answer.
3. Use returned memory only when it is relevant, permitted, and supported by citations.
4. After the task, write back a short summary of what happened and what future agents should remember.
5. Never store raw credentials, secrets, or private chain-of-thought. Store user-visible decisions, outputs, citations, and useful summaries.

// Before work
const memory = await parcle.memory.search({
  query: userRequest,
  tag: { project, user, thread }
});

// After work
await parcle.memory.messages({
  messages: [
    { speaker: "user", content: userRequest },
    { speaker: "agent", content: finalAnswerSummary }
  ],
  tag: { project, user, thread, task: taskId }
});

Authorization: Bearer <COMPANY_API_KEY>

What you can do#

Content is processed in the background and becomes searchable shortly after it is submitted, so nothing blocks while it is indexed.

Record a conversation#

Add the messages from an agent's work. Tag them so retrieval can be scoped later by team, project, or any dimension that makes sense.

{
  "messages": [
    { "speaker": "rep", "content": "We're moving Q3 pricing review to next week." },
    { "speaker": "agent", "content": "Noted: Q3 pricing review rescheduled to next week." }
  ],
  "tag": { "team": "sales", "topic": "pricing" }
}

{ "id": "mem_…", "status": "accepted" }

Add a document#

Add a file's content directly. A broad range of document and data formats are supported.

{ "filename": "q3-plan.md", "tag": { "team": "sales" } }

The file's content is ingested and indexed, then available to any agent permitted to see it.

Search memory#

Ask a question in plain language. The brain returns a grounded answer, composed from the relevant memory rather than matching snippets, with citations and a confidence signal. Use tag to scope the search to a slice of memory, or omit it to search everything the agent can access.

{ "query": "What did sales decide about Q3 pricing?", "tag": { "team": "sales" } }

{
  "answer": "Sales rescheduled the Q3 pricing review to next week.",
  "confidence": 0.86,
  "citations": [
    { "type": "conversation", "id": "mem_…" },
    { "type": "document", "id": "doc_…" }
  ]
}

• answer: a natural-language answer grounded in shared memory.
• confidence: how strongly the answer is supported by the retrieved memory.
• citations: the exact sources the answer drew on.

Manage memory#

Memory isn't write-only. Agents and admins can see what the brain holds and curate it, so you can always answer "what does the brain know, and where did it come from?" and keep it accurate.

• Browse: list what's in memory, filtered by tag, source, or time, to see exactly what the brain holds.
• Inspect: open a specific item to view its description, timeline, and where it came from.
• Update: correct or re-tag an item so memory stays accurate and well-grouped.
• Remove: permanently delete by item, source, or tag. Removal is permanent; to re-add content, ingest it again.

{ "tag": { "team": "sales" }, "since": "last quarter" }

{
  "items": [
    {
      "id": "mem_…",
      "type": "event",
      "summary": "Q3 pricing review rescheduled to next week",
      "tag": { "team": "sales", "topic": "pricing" }
    }
  ]
}

Browsing returns memory as events, each a concise summary with the tags and origin behind it, so you can review, correct, or prune at the level the brain actually reasons over. Access to these operations is governed like everything else; see Security.