AI agent memory glossary

The vocabulary of agent memory, in plain terms — and where MemQL fits each one.

Agent memory

The information an AI agent keeps across time so it can act coherently — what it has seen, done, decided, and learned. Without it an agent is stateless and starts from zero on every turn.

Agent memory database

A database purpose-built to store and serve an agent’s memory — not a general store you bolt onto an agent, but one whose data model is the agent’s state and history. MemQL is an agent memory database: an append-only, time-series graph on PostgreSQL + TimescaleDB.

Memory layer for AI agents

The part of an agent stack responsible for persistence and recall — the substrate the rest of the system reads from and writes to. In the Agent = Model + Harness model, the memory layer is the harness’s most load-bearing pillar.

AI harness

The runtime around a model that turns one LLM call into an agent that completes tasks — the loop, tools, budgets, retries, stopping rules, and memory. See the full explainer: What is an AI harness?

Temporal memory for AI agents

Memory that is aware of when things happened, so an agent can ask “what changed since yesterday?” or weight recent events more heavily. In MemQL every row is keyed by createdAt on a TimescaleDB hypertable, making time a first-class index rather than a metadata field.

Time-series graph (for AI)

A data model that is both a graph (records relate to one another) and a time-series (every record is timestamped and history is preserved). It lets an agent traverse relationships and reason over time in the same store — MemQL’s core shape.

Episodic vs. semantic memory

Episodic memory is the raw log of what happened (this step ran, this tool returned that). Semantic memory is what’s generally true, distilled from many episodes (“this approach usually fails here”). MemQL’s consolidation turns episodic memory into semantic memory on a schedule.

recall()

MemQL’s memory retrieval that blends recency and relevance in a single query — pgvector similarity combined with exponential time-decay over createdAt, scored server-side. It’s how an agent surfaces “the right memories,” not just the most similar ones.

Append-only (a.k.a. immutable) agent memory

A memory store where records are never edited or deleted in place — you only append new versions. Because the past is never overwritten, the full history and provenance survive, which is what makes an agent’s actions inspectable and replayable. People often search for this as “immutable agent memory”; MemQL is append-only, which gives that same guarantee.

Agentic memory layer

Shorthand for the memory layer specifically designed for autonomous agents — durable, time-aware, queryable, and aware of an agent’s working state (plans and steps), not just a passive cache of text.