Persist a piece of information to long-term memory so it can be recalled in future sessions. Use this whenever the user states a preference, makes a decision, or shares context that should survive beyond the current conversation.

When to call: after learning the user's tech stack, coding conventions, project constraints, architectural decisions, or personal preferences.

Returns a confirmation message with the stored content preview.

Examples of good memories:

• 'User prefers TypeScript strict mode with no implicit any'

• 'Database: PostgreSQL 16 with pgvector on Railway, connection via asyncpg'

• 'Never use any() type in this codebase — team policy'

• 'Deployed on 2024-03-15: migrated auth from JWT to Supabase sessions'

Search long-term memory using semantic similarity and return the most relevant stored memories ranked by a confidence score (weighted combination of similarity, recency, and access frequency).

When to call: before starting a task, when the user references something from a past session, or when you need project-specific context.

Returns a ranked list of memories with their confidence score and type. Returns an empty result if no memories exceed the similarity threshold.

Retrieve the most relevant memories for the current task and return them as a single formatted context block, ready to prepend to a prompt or include in a system message.

When to call: at the start of a complex or multi-step task where relevant project history, constraints, or preferences may exist in memory. Prefer this over recall when you want a single ready-to-use block rather than a list of individual memories.

Returns a formatted text block summarising the relevant memories, and a count of how many memories were used. Returns empty if none are relevant.

Search for memories matching a query and permanently delete the top matches. Use this to remove outdated, incorrect, or superseded information from memory.

When to call: when the user explicitly asks to forget something, or when you detect that a stored memory is no longer accurate (e.g. a dependency was upgraded, a decision was reversed, a team member left).

Internally performs a high-threshold semantic search (0.7) to find close matches, then deletes up to 3 results. Returns a list of deleted memories or a message if nothing matched.