Menu
Recurring workflows to deepen and enrich Hermes' memory systems (L1, L2, L3, OBn).
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Guides Cloudflare One Zero Trust and SASE work across Access, Gateway, WARP, Tunnel, Cloudflare WAN, DLP, CASB, device posture, and identity. Use when designing, configuring, troubleshooting, or reviewing Cloudflare One deployments. Retrieval-first: use current Cloudflare docs/API schemas instead of embedded product docs.
Use when scheduling recurring polls, read-only monitors, or background jobs that must survive shell session boundaries.
Discover, register, and maintain Hermes skills from non-default paths (plugin directories, external repos, shared vaults). Use when skills are missing from `skills_list()`, when the user mentions plugin skills, when symlinking skills, or when reconciling duplicate/overlapping skills between the default tree and external sources.
Operate the Antigravity CLI (agy): plugins, auth, sandbox.
Tune HERMES_STREAM_RETRIES for mid-stream reconnect resilience.
Manage parallel Hermes sessions in Telegram topics.
| name | memory-deepening |
| description | Recurring workflows to deepen and enrich Hermes' memory systems (L1, L2, L3, OBn). |
| version | 1.0.0 |
| author | Hermes |
All deepening workflows target:
fact_storemcp_contextforge_memory_* + Chroma 8000http://localhost:8888 (NOT 9876 — that port is a legacy/stale daemon)mcp_open_brain_* + Chroma 8001Auxiliary: session history (session_search) and profile memory (memory).
Purpose: Retain session summary to Hindsight, train Holographic trust scores, and capture key takeaways to Honcho + OB1.
Script: ~/.hermes/scripts/session-end-drain.py
Purpose: Seed all 4 memory layers when a novel workflow (e.g., Greek chorus) emerges. Steps:
memory(target="memory", action="add", ...)curl -X POST http://localhost:8888/api/lessons -d '{"content": "..."}'mcp_open_brain_capture_thought(content="...")honcho_reasoning(query="...")Reference: See Greek Chorus Pattern for session-specific detail. Steps:
python ~/.hermes/scripts/session-end-drain.py
Purpose: Prune L1, deduplicate L3, distill Honcho (ContextForge), and verify OB1/OBn queryability.
Script: ~/.hermes/scripts/weekly-memory-maintenance.py
Cron: 0 9 * * 1 (Mondays at 9 AM)
Steps:
python ~/.hermes/scripts/weekly-memory-maintenance.py
Purpose: Synthesize canonical memory contents into OBn vault notes and cross-fleet shared brain.
Script: ~/.hermes/scripts/monthly-synthesis.py
Cron: 0 9 1 * * (1st of every month at 9 AM)
Steps:
python ~/.hermes/scripts/monthly-synthesis.py
Purpose: Capture "remember this" requests into all four canonical systems.
Skill: enrich-memory
Usage:
hermes skill run enrich-memory --content "..." --entity "..." --category "general" --tags "..."
Purpose: Lightweight background enrichment every 30 minutes.
Pattern: dispatch up to 3 delegate_task workers:
fact_store).hermes/state/ and update it each cycle.Purpose: Synchronize Beads (Dolt-backed issue tracker) with Hermes memory systems (Holographic, Hindsight, Honcho). Steps:
bd export --issues --interactions | mcp_hindsight_capture_lesson
Pitfalls:
lsof -i :<port> to identify and kill conflicting processes.Purpose: Strengthen under-represented OB1/OBn topics without broad memory churn. Use when a worker is assigned a bounded L3 deepening lane such as replay boundaries, idempotency keys, writer-of-record, post-write verification, or BRAIN_KEY rotation closeout. Steps:
curl -sS -m 5 http://127.0.0.1:8001/api/v2/heartbeat
mcp_open_brain_search_thoughts, using topic-specific queries and expected hit thresholds. If the user asks for exact phrase hits, record before/after counts separately from semantic match scores.mcp_open_brain_capture_thought; prefer stable shared namespaces such as task_id=fleet/memory or task_id=fleet/decision, and set source to the actual writer surface.threshold=0.7, the chosen task_id, and the chosen source.ob1_ob1_core), check count, then use /api/v2/.../collections/<id>/get with where_document: {"$contains": "<phrase>"}. MCP semantic top-N exact counts and Chroma direct containment counts are different evidence types; report both if they diverge.See references/obn-gap-fill-exact-phrase-verification-20260619.md for a copy-paste verifier and reporting convention.
The three scripts under ~/.hermes/scripts/ were rewritten on 2026-06-18 to use real interfaces after the original versions imported non-existent hermes_tools functions and crashed. The actual hermes_tools module (available in execute_code context) only exports: read_file, write_file, search_files, patch, terminal, json_parse, shell_quote, retry.
Do NOT import hindsight_retain, fact_feedback, session_search, memory, hindsight_recall, fact_store, mcp_contextforge_memory_query, or mcp_contextforge_memory_delete_batch from hermes_tools — these functions do not exist there. They are Hermes agent runtime tools, not Python module exports.
The scripts now use:
sqlite3 direct access to ~/.hermes/memory_store.db (table facts, columns: fact_id, content, category, trust_score, helpful_count, deleted_at)urllib.request HTTP calls to localhost:8888 (gracefully skips if not running)subprocess calling ~/.hermes/bin/memory-prune.shmemory.memory_char_limit in ~/.hermes/config.yaml (currently 6000). Do NOT assume 2200 — that was a hardcoded default in memory-prune.sh that was fixed on 2026-06-18.memory-sop L4 quota section.trust_score by 0.05 per "helpful" call. Without regular runs, scores stay at the 0.5 floor. Run session-end-drain.py at every session end to train scores progressively.python3 processes MUST NOT import from hermes_tools. Use stdlib (sqlite3, urllib, subprocess, json) or installed packages instead. hermes_tools is only available inside execute_code and exports a limited set.