Skip to main content

Memory and Session

Session history and workspace memory are related but separate:

  • Session is persisted conversation history and the model-visible context projection for a single conversation.
  • Workspace memory is a developer convention: files like MEMORY.md, TASKS.md, or notes/*.md that the agent reads and updates through the bash tool.

The harness still loads MEMORY.md into the system prompt when the file exists. It no longer exposes separate memory or task-list tools. When a workspace has harness.enabled unset or true, the harness also adds short MEMORY/TASKS workflow instructions by default; the agent then manages memory and task files as ordinary markdown files in the mounted workspace.

Mental Model​

Workspace Sharing​

Workspaces are account-scoped records. Any agent or conversation that references the same workspaceId sees the same files:

Create a workspace, then reference it from the agent:

// POST /accounts/me/workspaces
{ "name": "notes", "config": { "storage": { "provider": "s3" }, "harness": { "enabled": true } } }

// agent config
{
  "sandbox": "sb_default",
  "workspaces": [{ "name": "notes", "workspaceId": "ws_notes" }]
}

Agents can expose multiple named workspaces. The first entry is the default when a tool call omits the optional workspace argument:

{
  "sandbox": "sb_default",
  "workspaces": [
    { "name": "personal", "workspaceId": "ws_personal" },
    { "name": "team", "workspaceId": "ws_team", "sandbox": "sb_locked_down" },
    { "name": "docs", "workspaceId": "ws_docs", "sandbox": null }
  ]
}

Runtime Behavior​

Session owns the runtime path:

  • claim() deduplicates an inbound event in ProcessedEvents.
  • acquireConversationLease() serializes work per conversation.
  • appendIngressEvents() persists incoming user, assistant, tool, and persisted system messages.
  • createTurnContext() loads conversation entries, builds system prompt parts, runs compaction when configured, and prunes model-visible messages.
  • workspaceBindings() resolves account-scoped workspace and sandbox records, applies per-workspace sandbox overrides, and hashes accountId:workspaceId with normalizeFilesystemNamespace().
  • filesystemNamespace() returns the default workspace namespace for existing single-workspace callers.

The namespace helper is in functions/_shared/runtime-keys.ts. The config interface and validation live in functions/_shared/storage/agent-config.ts.

Configure It​

Create a workspace with automatic MEMORY.md loading and default MEMORY/TASKS harness instructions:

{
  "name": "notes",
  "config": {
    "storage": { "provider": "s3" },
    "harness": { "enabled": true }
  }
}

Disable only the MEMORY/TASKS harness instructions while still loading an existing MEMORY.md:

{
  "name": "notes",
  "config": {
    "storage": { "provider": "s3" },
    "harness": { "enabled": false }
  }
}

Remove a workspace reference from the agent config to disable that workspace's mounted tools and prompt-time memory loading. Set workspaces[].sandbox: null when the agent should keep read-only read/glob access through S3 but must not mount or mutate files.

Session Context Management​

Session history is managed before each model turn:

  • Pruning is enabled by default unless session.pruning.enabled is false. It removes older reasoning/tool-call clutter from the model-visible context without changing persisted history.
  • Compaction is disabled by default unless session.compaction.enabled is true. When enabled, it uses the selected agent model to summarize older history once the serialized context exceeds session.compaction.maxContextLength.
  • Compaction persists a system summary, keeps the latest user message active, and includes prior compaction summaries when compacting again.