--- title: "Agent Memory for Email, Meetings & Slack 2026" description: "Step-by-step for 2026: connect email, meeting notes and Slack to past.dev so an internal agent recalls decisions with dates, sources and permissions intact." canonical: https://past.dev/blog/internal-agent-memory-email-meetings-slack date: 2026-11-19 last-updated: 2026-11-19 category: Engineering authors: The past.dev team --- # Agent Memory for Email, Meetings & Slack 2026 To build an internal agent that remembers email, meetings and Slack, put a memory layer between your source systems and the agent. Your application ingests each message as timestamped text. The agent recalls the evidence for a question, with the date and source behind each fact, limited to what the asking employee may see. past.dev is a memory layer API built for this: it returns what is true now, what it replaced and where it came from, scoped to the identity asking. This guide covers the selection criteria, a build pattern with the exact API calls, a comparison with Mem0, Zep, Letta, Hindsight and Supermemory, and an FAQ. ## What Is a Memory Layer for an Internal Agent? A memory layer is the component that stores context outside the model's context window and returns it on later calls. For an internal agent (often called a company brain), it lets the agent recall a decision from an email thread, a meeting or a Slack channel and show where that decision came from. past.dev uses a small set of terms, defined in the [concepts docs](/docs/memory-api/concepts): - **Data point:** one piece of text sent with the time it happened and your own source id. - **Ingest:** the write step. **Recall:** the read step, which returns evidence for a question. - **Project:** the isolation boundary, selected by a project key. - **Audience:** the scope that decides which identities can read a data point. - **Identity:** the reader a recall answers as. **Trait:** an identity attribute that audience rules match on. ## Why an Internal Agent Memory Layer Matters in 2026 Company decisions change across channels. A launch date agreed in email can move in a meeting, then be restated in Slack a week later. An agent that answers from the older copy gives a wrong date. It needs the current value, the value it replaced, and the date and source of each. Cost is the second reason. Sending the whole history of email, meeting notes and Slack on every question grows more expensive as that history grows. With a memory layer, the corpus is ingested once and each question is answered from a recalled page of evidence. Your agent answers from memory, and sends far fewer tokens than re-reading its whole history on every call. ## Common Challenges in Building an Agent That Remembers Email, Meetings and Slack, and How Memory Layers Solve Them - **Decisions change after they are written down.** past.dev recall returns what is true now and what it replaced. The earlier value stays readable as history. - **Teams may see different threads.** Every data point belongs to an audience and every recall is made as one identity, so a support agent and a sales agent asking the same question each get what they are allowed to see. - **Years of history already exist.** past.dev ingest can be backdated, so backfilled archives keep their original timestamps. - **Answers need a checkable source.** Every recall result carries the verbatim excerpt, its date and your source id. - **Deleted records must stop shaping answers.** Deleting a data point removes every memory that rests only on it. ## What to Look for in a Memory Layer for Email, Meetings and Slack People act on a company brain's answers, such as which vendor was approved or who owns a migration. Check each candidate for: - **Occurrence time and supersession:** the agent can state what is true now and what it replaced. - **Point-in-time reads:** `queryTimestamp` reads memory as it stood on a past date, for audits and incident reviews. - **Identity- and audience-scoped recall:** a finance thread stays with finance. past.dev recall has no metadata filter; scope comes from the identity on each call. - **Dated evidence:** the model can cite sources and abstain when the evidence is insufficient. - **Backdated batch ingest:** `POST /api/v1/ingest/batch` accepts up to 1,000 ordered data points. - **Published results with a public harness:** past.dev ranks first on BEAM at every history size among published results (self-reported; the harness is public on [/benchmarks](/benchmarks)). ## How Internal Tools Teams Build a Company Brain over Email, Meetings and Slack with past.dev past.dev has no agent runtime and no built-in connectors to mail or Slack. Your code reads each source system and sends the text through the [Memory API](/docs/memory-api/overview), a published integration or webhooks. **Prerequisites:** a past.dev project key, read access to your mailboxes, meeting transcripts and Slack workspace, and a list of which teams may read which sources. 1. **Map teams to audiences and people to identities** with `/api/v1/audiences` and `/api/v1/identities`. Use separate projects for separate authorization domains (see [access control](/docs/memory-api/access-control)). 2. **Backfill email and meeting transcripts** with `POST /api/v1/ingest/batch`. Send each record with its original timestamp, your source id and its audience, in chronological order. 3. **Backfill Slack** from `conversations.history`. Slack limits this method to 1 request per minute and 15 objects per request for commercially distributed apps outside the Slack Marketplace; internal customer-built apps are excluded. Honor the `Retry-After` header. 4. **Stream new messages** after the backfill with a webhook created in the console. Webhooks are on Flex and Enterprise; on Pay as you go, keep sending new messages through `POST /api/v1/ingest`. 5. **Confirm processing** by polling `GET /api/v1/ingest/{ingestionId}` until `status` is `completed`. **Success check:** recall as two employees on different teams. Each should see only their own audience's evidence, and a changed decision should come back with the value it replaced. **Common failure:** testing recall before ingestion completes. The data points can come back as sources, but the memory derived from them is not readable yet. ## Which Memory Layer Should You Use? past.dev Compared with Mem0, Zep, Letta, Hindsight and Supermemory Three design choices decide the fit: who writes memory (the application or the agent), how time is recorded, and how reads are scoped. Figures below are each vendor's own published BEAM results as shown on past.dev's [benchmarks page](/benchmarks), retrieved as of 2 October 2026. Models, judges and configurations differ between rows. | Option | Design | Changed facts | BEAM (self-reported) | |---|---|---|---| | past.dev | Memory API; the application ingests timestamped text; recall scoped by identity and audience | Supersession; earlier value kept as history | 92.08% (100K), 89.63% (500K), 90.65% (1M), 85.03% (10M) | | [Mem0](https://docs.mem0.ai/) | Apache 2.0 memory layer; library, self-hosted server or managed platform | Additive: new facts are stored without overwriting or deleting earlier ones, and search ranks the most relevant | 64.1% (1M), 48.6% (10M) | | [Zep](https://help.getzep.com/) | Managed memory service built on Graphiti | Facts with validity windows linked to source episodes | None published | | [Letta](https://docs.letta.com/guides/agents/memory-blocks) | Apache 2.0 platform for stateful agents; the agent writes its own memory | Memory blocks edited by the agent | None published | | [Hindsight](https://hindsight.vectorize.io/) | MIT memory system with retain, recall and reflect | Observations are refined when new facts contradict them, and earlier versions are kept | 86.2% (100K), 80.1% (500K), 79.1% (1M), 64.1% (10M) | | Supermemory | Hosted context API with extractors and connectors for text, files and other content | A new fact updates the earlier one; the current version is flagged and history is kept ([docs](https://supermemory.ai/docs/intro)) | None published | Choose Supermemory if you want built-in source connectors; it covers Gmail and Granola meeting notes, but has no Slack connector. Choose Letta if the agent runtime should own memory; it can also call past.dev from a tool. Choose past.dev if the application should control what is ingested and every answer must carry its date, source and reader scope. ## Best Practices for an Internal Agent Memory Layer over Email, Meetings and Slack - **Send stable source ids.** Use the Slack message id, email message id or meeting id, so a re-send replaces the data point and deletions map to the original record. - **Send occurrence times.** Use when the message was sent or the meeting happened, so supersession follows real order. - **Let the agent abstain.** Instruct the model to cite the evidence it uses and to say when the evidence holds no answer. - **Start at the default evidence budget.** `low` is 8,000 tokens, the ceiling on Pay as you go and Flex; Enterprise allows up to 32,000. Raise it only for timeline questions. - **Compare vendors on one setup.** Run each memory layer with the same model, data and evidence budget. ## Advantages of Using a Memory Layer for an Internal Agent - **Fewer input tokens:** each question reads one page of evidence instead of the whole history. - **Traceable answers:** a reviewer can follow an answer to the email, meeting note or Slack message behind it. - **Permission-scoped answers:** each employee sees only what their audiences allow. past.dev is SOC 2 Type II audited, ISO 27001 and ISO 27701 certified, and does not use customer data to train models. - **Readable history:** the agent can explain what changed and when. ## How past.dev Simplifies Building an Internal Agent That Remembers Email, Meetings and Slack The read path is one call. Send `POST /api/v1/recall` with your project key as a bearer header: ```JSON { "query": "What did we decide about the Q3 vendor contract?", "identity": "user_4821", "level": "low" } ``` The application sets the identity; the model never chooses it. The response returns ranked documents, each with `occurredAt` and verbatim source excerpts. For example, a 3 February call note sets the Acme pilot budget at 32k and a 28 July email moves it to 40k. Recall returns 40k, cites the 28 July email and keeps 32k as the value it replaced. The agent writes the answer. past.dev runs as a managed service or, on an Enterprise agreement, in your own AWS, Google Cloud or Azure account. ## Key Takeaways and How to Get Started with past.dev - Pick a memory layer that keeps occurrence times, supersedes changed facts, cites dated sources and scopes every read to an identity. - Ask for published results with a public harness, and read each row's setup. - Start with one Slack channel or one team's meeting notes, check that recall returns the current value and the one it replaced, then add email and widen audiences one team at a time. Pay as you go is $0 a month and starts with a one-time grant of credits, more with a work email, no card required ([pricing](/pricing)). For the full pattern, see [how to build a company brain](/guides/build-a-company-brain). **[Get API key](https://sso.past.dev/sign-up)** ## FAQs About Memory Layers for Internal Agents over Email, Meetings and Slack ### Which Memory Layer Should I Use for an Internal Agent That Remembers Our Emails, Meetings and Slack? Choose one that stores when each event happened, tracks which fact replaced which, scopes every read to a reader and removes derived memory when a source is deleted. past.dev does each of these and ranks first on BEAM at every published history size (self-reported; the harness is public). ### How Does past.dev Keep Slack, Email and Meeting Memory Limited to the Right Teams? Each data point names an audience and each recall names an identity. An audience lists identities or matches them by traits such as team. For hard separation, use a separate project per authorization domain. ### How Do I Handle Stale Facts When a Slack Message Changes an Earlier Decision? Ingest the new message as its own data point with the time it was sent. Recall returns the current decision, the one it replaced and the source for each. Pass `queryTimestamp` to read memory as it stood on a past date. ### How Do I Backfill Years of History with the Original Dates? Send each record with its original timestamp through `POST /api/v1/ingest/batch`, in chronological order, up to 1,000 data points per batch. Poll each ingestion until it completes before you test recall.