A featherweight memory supplement for Hermes that lives in your Obsidian vault. Zero dependencies. Zero cloud. Zero overhead.
Built-in memory runs out of tokens. Sessions reset. Context dies. This patches those exact gaps — and nothing more.
Three fundamental gaps remain in real-world use. This skill was built to close them — and nothing else.
Built-in memory caps at roughly 800 tokens — about one session's worth of substantive work. Start a second session and earlier context is already being squeezed out. It's not a flaw; it's an architectural limit. But it means yesterday's decisions compete with today's for the same finite space.
Every fresh /new or /reset is a blank slate. You have to re-explain who you are, what you were working on, what was decided. Session search helps — but it's manual, slow, and requires you to already know what to search for.
Today's built-in memory is invisible. It lives in opaque vectors and compression artifacts. You can't read what Hermes remembered. You can't correct it before it's used. If it captured the wrong thing from a session, you'll find out the hard way — when the next session starts with the wrong mental model already loaded.
It doesn't touch, override, or duplicate Hermes' built-in memory. It lives beside it — in plain Markdown files your agent reads and you can review.
No vector database. No embedding pipeline. No cloud relay. No background daemon. The entire system is a handful of Python scripts and Markdown files that sit in your Obsidian vault. Less than 26 KB downloaded. Zero install friction. Nothing to configure beyond pointing it at your vault.
Every captured session is a Markdown file you can open, read, and edit. Spot something Hermes recorded inaccurately? Fix it before the nightly summary runs. Want to add context? Write it in. The memory is yours — not a black box. This is the single biggest improvement over opaque vector stores.
Instead of splitting Hermes' attention between answering your question and managing memory, this uses Hermes' plugin hook system. pre_llm_call captures your prompt. post_llm_call captures the response. Both fire automatically — fire-and-forget observers that can never slow down or crash the main process. The agent never thinks about memory. It just works.
The nightly summary doesn't run on raw session data alone. You can review the day's exchanges, annotate decisions, correct errors, or flag what matters most — and only then does the cron job compress it. This turns memory from a passive log into an active collaboration between you and your agent. Most memory systems store things about you. This one stores things with you.
Purpose-built to fix the three gaps — not a kitchen-sink knowledge management platform.
CAPTURE Auto-Capture Every Turn |
Plugin hooks log your prompt and Hermes' response to the daily session file automatically, on every turn. No manual effort. No forgetting to save. No impact on what Hermes is doing — the hooks are observers, not participants. |
COMPRESS Nightly Summaries |
A cron job runs after midnight, reads the day's session file, and compresses only Hermes' responses into a structured summary — decisions reached, fixes applied, errors encountered, tasks left open. The raw log is preserved; the summary is what travels forward. |
RECOVER Context Injection at Session Start |
Every new session starts with the last 3 nightly summaries already injected into context. Hermes picks up exactly where things left off — without you re-explaining, without searching old sessions, without starting from scratch. |
TRACK Task Cross-Check |
Every nightly summary automatically compares newly identified pending tasks against your existing Todo Index. Nothing slips through and duplicates get caught before they pile up. |
PRIVATE Local-First, Always |
Zero external dependencies. No data leaves your Obsidian vault. No telemetry. No cloud. Everything runs locally on your machine or Hermes' terminal container — the same principle behind every ContextInit product. |
VERIFIED Test-Driven |
Built with TDD: every component has a test spec written before the implementation. The full suite is 42 tests, zero failures. Not because we got lucky — because each script was verified before the next one was written. |
From zero to working memory in under a minute. No config files to edit. No API keys to manage.
Unzip the package into your vault's Hermes_Agent_Folder. Run one command to register the cron job. That's the entire setup.
~ 15 secondsPlugin hooks record every turn in real-time. Your prompts stay verbatim. Responses are logged alongside. Nothing to configure, nothing to remember to do.
Zero manual effortThe last 3 summaries load automatically at session start. Hermes remembers. You don't repeat yourself. The context that was gone is simply there.
Seamless continuityv1.0.0 · MIT License · Works with any Hermes provider
↓ Free Download 26 KB zip · Full source included