Skip to content

tablestakesllc/meridian

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Meridian — Agent Context and Memory System

Version: 4.6
Publisher: TableStakes LLC
License: Apache 2.0

The Meridian Agent Context and Memory System is a file-native knowledge grounding system for AI coding agents. It has no server, no vector database, and no external API dependencies. Everything lives in the repository alongside your code.


The Problem

AI coding agents are stateless. Every session starts from scratch. The agent that debugged your pgBouncer issue last week has no memory of it today. It will rediscover the same gotchas, re-debate the same design decisions, and miss the same project conventions — every time.

Large tool vendors solve this with proprietary memory layers baked into their products. Meridian solves it with markdown files, a structured index, and an unbiased Python script.


How It Works

Meridian maintains a set of named memory blocks — discrete pieces of agent-curated domain knowledge about your codebase, stored as annotated sections in a plain markdown wiki.

Each block has a weight that changes over time based on how often it is loaded and used.

At the end of every session, a lightweight script updates block weights and regenerates a hot-load file (L1_CONTEXT.md). At the start of the next session, the agent reads that file before doing anything else. Frequently used knowledge stays in the hot-load file. Stale knowledge fades to on-demand retrieval. Nothing is ever deleted.

Three tiers:

Tier Weight Behavior
L1 ≥ 8.0 Auto-loaded every session
L2 2.0 – 7.9 Loaded on demand via skill trigger keywords
Archive < 2.0 Retained; recovers automatically when used again

Skills are domain-specific context files loaded in three ways: trigger keywords in the session, domain area, or a file path referenced directly from a curated memory block. They route the agent to the right knowledge for the work at hand without loading everything at once.


What's in the Repo

  • BOOTSTRAP.md: A step-by-step setup guide prompt for any new repository. Give it to a fresh agent and follow along. No prior Meridian knowledge required.
  • SHAKEDOWN.md: A 9-phase post-deployment verification prompt. Run it once after setup and after any major version upgrade to confirm the full installation is functioning correctly.
  • SEED_HISTORY.md: A structured prompt that analyzes a repository's git history to generate an initial set of memory blocks. Run it once on a new codebase to give Meridian a grounded starting point instead of a blank slate.
  • demo/: A reproducible test methodology and the prompts used to compare a Meridian-equipped agent against a baseline agent on an identical task.

The weight_memory.py script is defined in full in Appendix C of BOOTSTRAP.md. An agent can rebuild it from scratch on any repository using that specification.


Getting Started

  1. Copy BOOTSTRAP.md into the prompt of a fresh agent to instance Meridian.
  2. Optionally, copy SEED_HISTORY.md into the prompt of the same agent to populate your initial memory blocks from git history.
  3. Open a clean agent and run SESSION_INIT.md to begin your first session.

Total setup time on a new repository: 30–60 minutes.


Design Principles

File-native. Every Meridian artifact is a plain text file committed to your repository. No lock-in, no accounts, no services to run.

Tool-agnostic. Meridian works with any AI coding agent that can read files. VS Code Copilot, Cursor, Claude, a raw API call — if the agent can read a markdown file, it can use Meridian.

Earned, not configured. Weights change based on actual session usage, not manual tuning. Knowledge that gets used regularly rises to L1 automatically. Knowledge that becomes irrelevant fades without being deleted.

Pick what fits. Use the bootstrap and skip the weight system. Use the skill routing and skip the session lifecycle. Use the seed prompt to create any part or all of Meridian. There is no required minimum. All parts of Meridian work on their own.


Versions

Version Changes
4.3 Initial public release. Bootstrap, seed history, weight system, skill routing, session lifecycle.
4.4 Index-driven conflict detection for skill authoring. BAD/GOOD negative anchor rule for skills.
4.5 Hard checklist format for SESSION_INIT.md. Thin pointer pattern for agent auto-load file. weight_memory.py moved into the memory skill. meta.meridian_version in weights.default.yaml (L1 header auto-populated). Agent-agnostic skills path (.agents/skills/).
4.6 Maintenance counter with --maintenance-reset / --maintenance-defer. File-glob auto-dispatch (@{u}..HEAD). --check-defaults threshold diff. Dry-run L1 preview. Depends enforcement. --export-defaults and --calibrate. Auto-generated File→Skill routing table in L1_CONTEXT.md. Merge driver bootstrap in --seed. SHAKEDOWN.md post-deployment verification prompt. Pre-approved commands section in BOOTSTRAP.md.

Testing

In two double-blind trials, Meridian shaped the context, grounded the memory, and guided the agent to fully complete the task. The baseline agent (same model, no Meridian) didn't adhere to the prompt as well and took longer to assess the work. See demo/TEST_PLAN.md for the full methodology.


License

Copyright 2026 TableStakes LLC. Licensed under the Apache License, Version 2.0.
See LICENSE for the full text.

About

File-native agent context and memory system for AI coding agents

Topics

Resources

License

Stars

Watchers

Forks

Contributors