memora

name	memora
description	Query and manage the Hermes Memora long-term memory worker (Cloudflare Workers). Semantic search, fact CRUD, stats, and nightly indexing.
version	2.0.0
author	Memora Contributors
license	MIT
metadata	{"hermes":{"tags":["memory","rag","search","knowledge-base","cloudflare","hermes-plugin","memora"],"category":"research","related_skills":[]}}

Memora: Persistent Semantic Memory

Connects Hermes to the persistent RAG worker for cross-session long-term memory. The RAG worker stores facts in D1 and embeddings in Vectorize, enabling semantic search across indexed memories.

When This Skill Activates

Use when the user:

• Asks to recall something from past sessions or long-term memory
• Wants to search the knowledge base
• Needs to add, update, or delete a persistent fact
• Asks for memory stats or to export the memory database
• Wants to run nightly maintenance (memora-nightly) or generate weekly digests (memora-weekly)

Tools

Memora provides the following tools:

• memora_search: Semantic search across all indexed memories. Returns ranked results by relevance. Use for conceptual queries.
  • Parameters: query (string), top_k (integer, default 10)
• memora_think: Synthesize an answer across retrieved memories with citations and gap analysis. Use this when the user asks a question that requires connecting multiple facts.
  • Parameters: query (string), top_k (integer, default 10), scope (string, optional)
• memora_list: List facts with SQL filters. Good for browsing or finding facts by exact category.
  • Parameters: category (string), search (string), limit (integer), offset (integer)
• memora_stats: Get a breakdown of facts by category and total count.
• memora_add: Persist a new fact to long-term memory.
  • Parameters: content (string), category (string, default "memory"), scope (string, "personal" or "company"), id (string, optional)
• memora_update: Update an existing fact by ID.
• memora_delete: Delete facts by ID.

Shared company rule files

If the company memory repo contains files at the root such as _brain-filing-rules.md or _output-rules.md, they are automatically loaded into your system prompt. Follow them when filing facts or formatting answers.

Core Workflows

1. Recall & Search

1. When the user asks a question about past context, call memora_search with the user's query.
2. If exact categories are known (e.g., business, projects), use memora_list.
3. Do not assume facts are missing if they were just added. Vector indexing has a slight latency.

2. Capture & Update

1. After learning something important (e.g., investor preferences, strategic pivots), DO NOT use the default memory tool. ALWAYS call memora_add directly to push to the RAG backend.
2. Categorize appropriately (e.g., user for preferences, business for strategy, project for execution). MUST be specific, avoid the default 'memory' bucket.
3. If a previous fact is contradicted, use memora_update to revise it rather than adding a duplicate.

3. Maintenance & CLI Scripts

The Memora Python package provides CLI scripts for workspace maintenance. Run these via the terminal tool when requested by the user or scheduled via cron:

• memora-nightly: Nightly brain indexer. Scans workspace files (mtime hashing), extracts entities for the wiki, and detects contradictory facts. Run via cd ~/hermes-workspace && memora-nightly.
• memora-weekly: Generates a weekly digest of memory activity. Run via cd ~/hermes-workspace && memora-weekly.
• memora-doctor: Health check. Reports worker reachability, pending vector sync, local queue depth, and company repo sync lag. Run via cd ~/hermes-workspace && memora-doctor.

Installing Memora for Other Agents

If the user asks how to set up Memora on a new agent or workspace, provide these steps:

1. Install Plugin:

   cd ~/.hermes/plugins/
   git clone https://github.com/fenestbuc/memora.git
   pip install -e memora/
   

2. Enable in config (~/.hermes/config.yaml):

   memory:
     provider: memora
   plugins:
     enabled:
       - memora
   

3. Configure Environment: Set RAG_WORKER_URL and RAG_AUTH_TOKEN.
4. Install Skill: Copy docs/SKILL.md from the repo into ~/.hermes/skills/memora/SKILL.md.

Troubleshooting & Constraints

• Trivial Messages: The plugin auto-filters short (< 15 chars) or trivial messages (e.g., "ok", "go ahead") from being ingested to prevent corpus pollution.
• Vector Index Latency: Facts added via memora_add may take a few seconds to appear in memora_search due to embedding generation. This is normal.
• Worker Configuration: The Cloudflare Workers RAG backend lives in rag-worker/ inside this repo. It is deployed separately from the Hermes plugin. Check health using memora_stats.