Events
An Event is an atomic interaction — a user message, an AI action, a sensor reading, or a reasoning step. Every piece of data enters Minns Memory Layer as an event. Events carry rich context including:- Who — the agent and session that produced the event
- What — the event type and payload (Action, Observation, Communication, Cognitive, Learning, or Context)
- Why — the active goals driving the interaction
- Where — the environmental state at the time of the event
Episodes
An Episode is a sequence of events forming a coherent task. Minns Memory Layer automatically groups events that share the same active goal into episodes. When a goal’sprogress field reaches 1.0, the episode is marked as complete and becomes a candidate for long-term memory.
You control episode boundaries by updating the
progress field on your goals. Set progress: 0.0 at the start and progress: 1.0 when the task is done.Goals
A Goal is the objective driving an episode. Goals are embedded in every event’scontext.active_goals array, which lets Minns Memory Layer understand that a Communication event and an Action event belong to the same task.
| Field | Type | Description |
|---|---|---|
id | u64 | Unique goal identifier |
description | String | Human-readable description (e.g., "book_movie") |
priority | f32 | Priority level (0.0 to 1.0) |
progress | f32 | Completion progress (0.0 to 1.0) |
deadline | u64 | null | Optional deadline timestamp |
subgoals | u64[] | IDs of child goals |
Memory
A Memory is the long-term record of a significant episode. Once an episode completes, Minns Memory Layer consolidates it into a memory that can be recalled later by matching the current context against past contexts. Memories are typed:- Episodic — a full episode record
- Working — short-term, high-relevance memory
- Semantic — distilled facts and knowledge
- Negative — records of failed approaches to avoid repeating mistakes
Claims
A Claim is an atomic fact extracted from event content — for example, “User prefers Sci-Fi movies” or “Seat H12 was booked successfully.” Claims are extracted automatically when you setenable_semantic: true on event submission. They power the semantic search system (POST /api/claims/search), which lets you query soft knowledge using natural language.
Strategies
A Strategy is a learned behavioral pattern extracted from successful episodes. If an agent repeatedly follows the same sequence of actions to achieve a goal, Minns Memory Layer distills that pattern into a reusable strategy. Strategies include:- Reasoning steps — the ordered sequence of decisions
- Quality score — how well the strategy performs
- Preconditions — when the strategy applies
- Action hints — what to do next