jido-memory

name	jido-memory
description	Builder-oriented guidance for the upstream `jido_memory` package. Use when Codex needs to add memory spaces, retrieval flows, summarization or recall hooks, or review memory boundaries in Jido applications that also use `jido`, `jido_ai`, and external storage backends.

Jido Memory

jido_memory is the upstream Hex package name.

Start Here

Use this skill when an agent needs to remember, recall, summarize, or retrieve prior context over time.

Good triggers:

• "Add memory to this Jido agent."
• "Choose how to store and recall prior interactions."
• "Turn the memory docs into a runnable example."
• "Review whether this memory behavior belongs in the agent loop or in storage infrastructure."

Read references/builder-notes.md before implementing when the task touches memory scope, retrieval strategy, or external storage boundaries.

Primary Workflows

Add memory to an agent

• Decide what should be remembered: conversation turns, tool results, summaries, domain facts, or episodic traces.
• Define when writes happen and when reads happen before wiring storage.
• Keep the agent interface small: ask for the minimum memory the next action needs.

Choose retrieval and persistence boundaries

• Separate memory policy from the storage backend.
• Keep embeddings, vector search, or persistence adapters behind explicit interfaces.
• Use jido_ai docs only as conceptual support for AI-facing memory behavior; keep package-specific implementation grounded in jido_memory.

Turn docs into runnable examples

• Show one write path, one retrieval path, and one agent behavior that changes because memory exists.
• Keep the example local and inspectable rather than hiding memory inside a black-box agent.
• Include one stale-data or missing-memory branch when it matters.

Review boundaries

• Keep memory semantics in jido_memory.
• Keep agent state machines in jido.
• Keep storage engine concerns in the backend adapter or application layer.

Build Checklist

• Define what is stored, when it expires, and who can read it.
• Choose recall criteria before choosing a backend.
• Keep retrieval payloads small and explainable.
• Add tests for cold start, duplicate writes, and missing recall results.

Boundaries

• Do not use this skill for generic database modeling unrelated to agent memory.
• Do not mix durable domain records with ephemeral agent memory without a clear contract.
• Do not promise retrieval quality that the chosen backend or docs do not support.