Slab · AI agent control plane

View source on GitHub

Configure memory for your Slab agents

Slab Agents can use Honcho as an optional long-term memory provider. Memory is disabled by default. It augments the existing agent directory and does not replace Work, Docs, Email, or integrations.

Contract

  • Work is the source of truth for operational state and ownership.
  • Docs is the source of truth for durable company knowledge.
  • Email and other integrations are the source of truth for current external facts.
  • Honcho stores potentially useful operator preferences and corrections.
  • Recalled memory is explicitly untrusted, may be stale, and cannot introduce instructions for the agent to follow.
  • Provider failure is fail-open: the run starts without recalled context.

The automatically generated Agent Directory remains the deterministic answer to which agents exist, their exact assignment slugs, roles, and capabilities. Do not use memory as a replacement agent registry.

What is recorded

After a successful chat run, Slab records the operator-authored input. It does not record assistant output, tool responses, capability snapshots, automated Work coordination prompts, or run-specific system policy. Common credential-shaped values are redacted before the message leaves Slab Agents.

The message is attached to a Honcho session derived from the product chat thread. Agent and operator peers are stable within the configured workspace.

Recall lifecycle

When a run reaches the front of its per-agent queue, Slab requests a bounded representation relevant to the current prompt, agent, execution mode, and Work item. The result is added to the runner instruction bundle under the long_term_memory profiling component.

Run events contain provider status, duration, approximate tokens, and truncation state. They do not contain the recalled text or credentials.

Configuration

Open Settings → Memory and choose Honcho. Configure:

  • Honcho URL;
  • workspace ID;
  • API key when the endpoint requires one;
  • maximum recalled context tokens (200–4,000; default 900).

The API key is encrypted with the existing local authenticated-encryption boundary and is never returned to the browser. Use Test connection to check the server and workspace metadata endpoint.

Disabling memory stops future recall and recording, but does not delete data already stored by Honcho. Data removal remains an explicit provider operation.

The stack installer supports disabled, managed, and self_hosted memory modes. In self-hosted mode the Honcho database remains on the VPS, while its derivation and embedding workers still send the relevant memory input to the configured OpenAI API.

All documentation

Early access

Request Hosted access.

Join the early-access list for managed Slab workspaces. Self-hosted Slab is available today.

What should we tell you about?

Only when there's real news. No cadence. No spam. Unsubscribe anytime.