Save and recall Memory

Memory keeps durable decisions, constraints, and facts. Each new memory is an independent atomic-memory Artifact; see Use Atomic Memory for interfaces, lifecycle and legacy API compatibility. PreparedContext selects relevant history for one request; it is temporary and does not create another Memory.

Save, recall, and correct

1. Complete the Quick Start and keep the resolved Scope consistent across sessions.
2. Explicitly ask the Agent to save the information. A direct remember_memory write does not require a model.
3. Search with search_memory, or inspect entries with list_memory_entries and get_memory_entry where the host exposes them. Preserve the exact ArtifactRefs they return; exact history is read as Atomic Memory revisions.
4. Replace content through the generic Artifact API or mark it forgotten through Atomic lifecycle. Content and history remain available for restoration. Replacement requires the content ETag; forgetting requires the exact reference and state_version.

Host tool names differ; see Connect Agents. The HTTP API exposes the complete request schemas and concurrency requirements.

FTS uses the same query terms for candidate retrieval and admission. Query normalization excludes common English function words in longer queries and recognizes standalone execution instructions accompanying a question, such as "Use only supplied context", "Do not call tools, read files, inspect old sessions, or delegate", and "Answer briefly and cite the artifact reference from the provided context. If no reliable context is present, say so." Those instructions cannot supply evidence for an unrelated fact. Explicitly quoted terms remain searchable, including function words used as identifiers such as "AND" "OR" precedence. Domain-specific directives and unrecognized wording remain searchable; this is a conservative lexical rule, not a general intent classifier. The default threshold requires 25% of all remaining distinct terms, with a minimum of two matches; one- or two-term queries require one match. Direct search and prepared context share this rule, including when the optional recall gate relaxes the threshold. Stored text and vector queries are unchanged, and matching words alone do not establish semantic relevance. When the entire query consists of execution instructions, its terms remain searchable so a caller can find that policy.

Automatic context and extraction

Recall hooks ask the Server for bounded PreparedContext. An empty result is valid when no relevant information is available. Returned history does not override current instructions or live project state.

Use Prepare standard context text to choose Memory, Experience, Profile, and Topic Memory sections, set their order and limits, and configure the total entry limit.

Capturing a prompt creates Source evidence. It does not guarantee Memory extraction. Enable a generation model and Source processing through model configuration; configure embeddings only when vector or hybrid search is needed. Review the host's prompt-capture switch before recording project input.

Topic Memory has separate processing and retrieval surfaces (search_topic_memory and get_topic_memory on MCP). It is not enabled by a basic explicit Memory write; inspect the Topic Memory settings and enabled runtime capabilities in Configuration.

Use Handoff to transfer the current task boundary, and Sources to preserve raw evidence.