Workspace
Memory
Account memory gives Swarm persistent knowledge across sessions. It preserves project conventions, private operational facts, and recovery guidance under explicit human approval and strict security boundaries.
Overview
Context that carries forward safely without leaking secrets.
In traditional LLM sessions, conversational context vanishes the moment a session closes. Re-explaining repo architecture, build nuances, or deployment environments wastes time and introduces drift.
Swarm addresses this through Account Memory: a durably stored, structured knowledge layer attached to your account and workspaces. Swarm can inspect active memory entries to guide runs, remember verified operational patterns, and recall project orientation—all protected by explicit permission gates and version concurrency.
Data model
Kinds and purposes keep facts distinguishable.
Memory entries are categorized by their epistemic trust level (Kind) and functional role (Purpose):
rule
Explicit user-approved operating rules and behavioral constraints. Swarm prioritizes approved rules during run orchestration.
orientation
High-level project routing and repository structure. Tells Swarm where subsystems live and what tasks belong to which workspace.
learned
Verified empirical facts and operational observations recorded during previous runs (such as host network configurations or diagnostic steps).
Entries also declare their operational Purpose:
preference: User interaction preferences and workflow habits.project_context: Domain terminology, service architecture, and repo layout.operational_context: Machine paths, non-secret configuration locations, and service ports.recovery: Safe, tested recovery procedures for known failure modes.
Safety & security
Mutations require explicit user consent and revision guards.
Memory persistence is strictly controlled to prevent prompt injection and data exfiltration:
permission gating
Remembering, editing, or forgetting a memory item requires an explicit user approval prompt detailing the intent and revision.
compare-and-swap (CAS)
Updates require the exact expected_revision token. Concurrent edits or stale writes fail closed without overwriting state.
no secret storage
Memory is designed for operational guidance, not secret management. Swarm enforces that credentials, API keys, and tokens are never saved into memory. Only configuration paths and credential locations may be referenced.
delegation scoping
Delegated child agents (Coder, Finder, Designer) do not gain direct memory tools or account selection. Only the primary orchestrator holds contextual memory.
Orientation
The Workspace Map routes work across linked repositories.
The Workspace Map is a special orientation document maintained in memory. It provides a lightweight catalog of your account's workspaces, their specific purposes, and routing keywords.
When your workflow spans multiple interconnected repositories (such as a backend API repo, a frontend web repo, and an infrastructure repo), the Workspace Map helps Swarm determine the canonical source of truth for any task without guessing.
Commands & interface
Inspect and manage memory directly.
You can view and manage account memory at any time in the Swarm Terminal UI or Desktop interface:
/memory
Opens the interactive memory manager in the terminal, displaying current rules, orientation, and learned entries with revision counts.
conversational requests
Ask Swarm in chat: "Remember that our staging database requires an SSH tunnel on port 5433" or "Show me what rules are currently saved in memory."