Agent memory

Three memory layers

Episodic notes and living digests

Conversations produce structured notes over bounded message windows. Notes record the topic, what happened, decisions, open questions, entities, and interaction signals. They retain pointers to the covered messages so users can inspect the source.

A living digest folds completed notes into a compact current state for long-running threads. A small language-model step synthesizes the configured 1–6 conversation starters from eligible digest context; provider failure falls back to deterministic digest text. “All sources” still applies the private destination-participant subset rule, while “public only” excludes private threads.

User preferences

Preferences belong to a user and are scoped separately for each agent. Examples include writing style, output format, language, and workflow habits. Inferred records expose confidence and the source conversation that last revised them.

The AI Preferences settings page lists every stored preference and lets the user edit or delete it. Agents may use only preferences belonging to a participant whose scope and permissions allow that use. Agent managers need separate view/manage preference permissions to see another principal's record, and those reads and writes are audited.

Agent semantic memories

Semantic memories are durable facts associated with one agent: learned domain facts, stable project context, or knowledge the agent deliberately records. They are managed through the agent's Memory capability, where authorized users can inspect and delete individual entries.

Semantic memory is not a hidden file browser and is not the raw transcript archive. The API returns typed memory records with source metadata, timestamps, and capability checks. Tags are normalized to lowercase and can narrow list and recall results.

Search and recall

• Folder-scoped document search never escapes the caller's authorized folder.
• Conversation-note search combines lexical and semantic ranking.
• Recall checks current notes, exact messages, then participant-scoped prior digests. Strong digest hits include provenance; moderate hits require confirmation; no hit is reported instead of guessed.
• Preference and semantic-memory recall use their own principals and capabilities.

Transparency and deletion

Memory must be inspectable at the same level where it is used: notes beside a conversation, preferences in the user's settings, and semantic memories on the agent. Generated notes cannot be reprocessed from the file preview because their source of truth is the conversation pipeline.

Deleting a note, preference, semantic memory, conversation, or agent follows that resource's retention and audit policy. No UI control implies a cross-layer purge.