---
title: "What is a memory API?"
description: "A memory API defined: the interface for writing records with timestamps, polling settlement, recalling with evidence, and deleting by source."
canonical: https://past.dev/glossary/memory-api
last-updated: 2026-09-02
---
# What is a memory API?

Source: https://past.dev/glossary/memory-api

A memory API is a service interface for storing and recalling an application's knowledge: raw records go in, structured and attributed answers come out. Instead of running its own extraction, entity linking, and temporal logic, an application sends raw records to the API and asks questions later. The interface typically covers four operations: writing records with timestamps, polling until processing settles, recalling with evidence, and deleting by source.

## What the interface covers

A memory API has four concerns, whatever the vendor: getting records in, knowing when they are usable, getting answers out, and removing what should no longer be there.

| Operation | What it does |
| --- | --- |
| Write | Accepts a raw record, its original timestamp, and a stable source id. The timestamp matters because records are often ingested long after the events they describe. |
| Settlement | Reports when asynchronous processing has finished for a given write, so callers know when recall reflects it. See [memory consolidation](/glossary/memory-consolidation). |
| Recall | Takes a question and returns ranked evidence with dates, sources, and a status value stating whether the evidence supports an answer. |
| Delete | Removes a source and the facts only that source supported, for corrections, retention policies, and erasure requests. |

## Why applications use one

- **One pipeline instead of many.** Coreference, [fact extraction](/glossary/fact-extraction), [entity resolution](/glossary/entity-resolution), and supersession run behind a single write call rather than inside every application.
- **Shared memory.** Several agents and channels write to and recall from the same store, so knowledge from email is available in chat.
- **Governance in one place.** Retention, deletion, and [audit trails](/glossary/audit-trail) attach to the store rather than to each agent.
- **Answers with evidence.** Recall returns dated, source-attributed results rather than bare text, which supports [grounding](/glossary/grounding) and abstention.

## A memory API in practice

A support agent finishes a call with Harborlane on March 12 and writes the transcript to the memory service with the call's timestamp and a source id. The application polls settlement until the record is processed. On October 7, before replying to a Harborlane ticket, the agent recalls the account's invoicing rules and receives the March 12 instruction with its date, its source, and a status value saying the evidence supports the answer.

When Harborlane later requests erasure of the call, deleting that source removes the facts only it supported, and subsequent recalls answer from what remains. past.dev implements this interface with four endpoints under `/api/v1` (ingest, settlement polling, recall, and per-source delete); the [quickstart](/docs/memory-api/quickstart) shows a first write and recall.

## Related concepts

- **[Memory consolidation](/glossary/memory-consolidation)**: the background processing settlement reports on.
- **[Recall](/glossary/recall)**: the read operation, ranked, dated, and attributed.
- **[Evidence status](/glossary/evidence-status)**: the label recall attaches to each answer.
- **[Cross-session memory](/glossary/cross-session-memory)**: the capability a memory API exists to provide.
- **Agent memory overview**: [what is agent memory](/what-is-agent-memory) places the API in the wider architecture.

## Frequently asked questions

### What does a memory API do?

It accepts raw records such as emails, transcripts, and notes, structures them into facts about entities over time, and answers later queries with ranked evidence that carries dates and sources.

### How is a memory API different from a vector database?

A vector database is infrastructure: the application embeds, stores, and queries chunks itself. A memory API is a service contract that handles extraction, entity resolution, and temporal tracking behind write, recall, and delete operations.

### How do I add memory to an AI agent?

Send each record the agent should remember to a memory service with its original timestamp, wait for processing to settle, then have the agent call recall with a question before answering. Delete by source when data must be removed.

## Related

- [What is memory consolidation?](https://past.dev/glossary/memory-consolidation)
- [What is recall?](https://past.dev/glossary/recall)
- [What is an evidence status?](https://past.dev/glossary/evidence-status)
- [Memory deletion guide](https://past.dev/guides/gdpr-memory-deletion)
- [Benchmark methodology](https://past.dev/benchmarks/methodology)