Fidelis memory workflow

SkillDocs & knowledge

Recall local records when a turn depends on prior work or decisions, or store and correct facts the user intends to retain. Requires a running Fidelis service; provides fidelis_recall, fidelis_store, fidelis_correct, fidelis_get, fidelis_recent, and fidelis_health.

Use Fidelis memory workflow in Claude, ChatGPT or Ahel Desktop

Free. Sign in, add Fidelis memory workflow and connect your AI. About a minute.

Also: Claude Code · Cursor · Codex

Then ask your AI: use the Fidelis memory workflow skill

Details

Instructions available. Your AI can read the instructions. Execution depends on the setup they require.

Add Ahel to your AI once: Claude, ChatGPT, Cursor, Claude Code or Codex. Then ask it to use this.

Fidelis memory workflowStart free

What this skill tells your AI

The instructions your AI receives, as published by hermes-labs-ai/fidelis in claude-plugin/skills/fidelis-memory/SKILL.md and read by Ahel’s review.

Use fidelis_recall when the current turn needs earlier evidence. Default fast recall uses local embeddings without a generative LLM. Use as_of for historical validity and mode: "thorough" for explicit hybrid retrieval.

Read the record ID and temporal status alongside the text. Superseded records remain visible as history. Quote the original wording and preserve constraints; retrieval does not authenticate a claim or prove completeness. Similarity is a ranking signal, not calibrated confidence. Empty results do not prove absence.

Use fidelis_get to fetch full text and correction links by ID. Use fidelis_recent to browse recent records or corrections. On errors, inspect fidelis_health; an unavailable or unloaded store is not an empty one. Ask the user to start the local service and Ollama when needed.

When the user intends a durable fact to be kept, call fidelis_store. To replace an earlier statement, call fidelis_correct with its ID and the replacement text. The original remains in history. Optional validity dates describe when the statement applies. Read the acknowledgement: queued is not stored, duplicate creates no new fact, and rejected writes were not accepted. Do not invent success.

Do not store secrets or content the user has excluded. Storage is local and not application-encrypted. Do not call recall for self-contained turns.

FIDELIS_PORT must match on the service and MCP process. Restart the client after upgrading to refresh its six-tool surface.

Signals

GitHub stars
23
Forks
3
Last commit
Oct 2026
Advanced
Item type
skill
Key
fidelis-memory
Source
github.com/hermes-labs-ai/fidelis