explaingit

duclamvan/hermes-memorykit

★ 0PythonAudience · developerComplexity · 4/5ActiveLicenseSetup · moderate

TLDR

Memory add-on for Hermes Agent that turns chat transcripts into notes, an entity graph, and ranked focus briefs so long-running agents stop forgetting facts.

Mindmap

mindmap
  root((hermes-memorykit))
    Inputs
      Chat transcripts
      LCM SQLite store
      Markdown notes
    Outputs
      Entity graph
      Focus brief
      Regression report
    Use Cases
      Persist agent memory
      Build focus brief
      Run nightly indexes
    Tech Stack
      Python
      SQLite
      Hermes Agent

Things people build with this

USE CASE 1

Give a Hermes Agent durable memory across compressions, cron runs, and tool calls.

USE CASE 2

Extract an entity graph of people, projects, topics, files, and sessions from Markdown and the LCM store.

USE CASE 3

Build a focus brief with citations that the agent reads before each task using RRF retrieval.

USE CASE 4

Run nightly maintenance and regression checks to catch memory drift in long-running agents.

Tech stack

PythonSQLiteHermes Agent

Getting it running

Difficulty · moderate Time to first run · 1h+

Requires an existing Hermes Agent install with the LCM SQLite store, plus a .env, config.yaml, and verify script pointed at your Hermes home and workspace.

MIT license, so you can use, modify, and redistribute the code commercially as long as you keep the copyright notice.

In plain English

Hermes MemoryKit is an add-on for users of an AI agent system called Hermes Agent. The problem it tries to fix is that long-running AI agents tend to forget things: after a long chat is compressed to save space, after a scheduled cron run finishes, or after a tool call replaces older context with newer context, important facts can disappear. This kit gives the agent a layered memory system so it stops losing track. The pipeline, as the README sketches it, goes from a raw chat transcript to durable notes, then to searchable Markdown documents, then to an entity graph that links people, projects, topics, files, and sessions, then to a router that ranks all those sources using a technique called RRF (Reciprocal Rank Fusion), then to a short focus brief with citations that the agent reads before each task, plus regression tests to catch memory drift and nightly maintenance to keep indexes fresh. The author reports a companion benchmark scoring 100 out of 100 with 27 retrieval checks passed. What you actually install is a Python package with scripts for each stage: entity graph extraction from Markdown and from a SQLite transcript store called LCM, a hybrid retrieval router, a focus-brief builder, a regression checker, a stack verifier, and a nightly maintenance job. There are also config templates for a .env file, a Hermes config.yaml, and example cron prompts. An optional Hermes plugin wrapper turns the scripts into native Hermes tools called memory_stack_status, memory_stack_route, memory_stack_focus_brief, and memory_stack_regress. Quick start is a git clone, a Python virtual environment, pip install -e with the dev extras, copying the example .env, and running a verify script that points at your Hermes home and workspace folders. The README also gives a minimal Hermes config that enables memory, sets the engine to LCM, and turns on compression with a context threshold of 0.70. A safety note warns you not to publish your raw LCM database, private notes, session IDs, or local paths. License is MIT.

Copy-paste prompts

Prompt 1
Walk me through the hermes-memorykit pipeline from raw transcript to focus brief and where each script fits.
Prompt 2
Show me the minimal Hermes config.yaml that enables memory, sets the engine to LCM, and turns on compression at 0.70.
Prompt 3
Explain how the RRF router combines entity graph, Markdown search, and LCM notes into a single ranked context.
Prompt 4
Help me wire the memory_stack_focus_brief and memory_stack_regress plugin tools into my existing Hermes setup.
Prompt 5
Set up the nightly maintenance job and the regression checker so I get an alert when memory drift exceeds a threshold.
Open on GitHub → Explain another repo

Generated 2026-05-22 · Model: sonnet-4-6 · Verify against the repo before relying on details.