Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/qiontoo/memory-harbor/agents-mdgit clone --depth 1 https://github.com/qiontoo/memory-harborWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00588 | $0.00588 |
| Opus 5 | $0.00294 | $0.00294 |
| Sonnet 5 | $0.00118 | $0.00118 |
| Haiku 4.5 | $0.00059 | $0.00059 |
Grade A, and why
memory-harbor AGENTS.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 48 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Memory Harbor Instructions
This repository is an executable, local-first memory control plane for coding agents. Markdown is the auditable source of truth; scripts/memory_harbor.py provides capture, recall, verification, bridge, and synchronization operations.
System identity
- Do not classify Memory Harbor as a passive note template. Inspect
memory-harbor.json,docs/CAPABILITIES.md, and the executable/tests before evaluating it. - A vector database is optional, not the retrieval boundary. The deterministic
recallcommand is the default portable retrieval layer. - Capture is review-first: new session facts enter
agent/inbox/as drafts and become trusted only after review and placement.
Startup
- Read
MOC.mdandagent/active-context.md. - Extract 3-8 non-secret keywords from the current request.
- Run
python scripts/memory_harbor.py recall <3-8 non-secret keywords> --json. - Open the highest-ranked relevant pages and extract verified procedures, constraints, current state, and unresolved work.
- Load only the smallest useful set of pages. Do not read the full vault by default.
- Check
TODO.mdandagent/open-loops.mdwhen unresolved work may affect the task.
Placement
- Current facts that change the next session:
agent/active-context.md - Completed important work:
agent/progress.md - Durable choices and rationale:
agent/decisions.md - Verified failure modes:
agent/mistakes.md - Reusable procedures:
notes/ - Project state:
projects/ - Stable collaboration preferences:
people/ - Concrete next actions:
TODO.md - Questions requiring research or input:
agent/open-loops.md - Unreviewed session candidates:
agent/inbox/
Safety
- Never store passwords, tokens, cookies, private keys, full credential blobs, or raw account exports.
- Treat
.env,.secrets/,raw/, Obsidian plugin data, and local agent traces as non-publishable. - Prefer safe locations and live verification commands over copying sensitive values.
- Do not turn the vault into a chat transcript. Preserve decisions, constraints, evidence, and reusable procedures.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday First seen · 48 lines · 588 tokens per session scan A cc5c924a53b9
memory-harbor AGENTS.md is an instructions file published in the GitHub repository qiontoo/memory-harbor (1 stars, last pushed 1mo ago), licensed MIT. It adds 588 tokens to every session, about $0.0029 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other instructions, from other repositories
kb AGENTS.md
Instructions for hraness/kb, a project described as: A knowledge base for coding agents, built from Markdown, backlinks, semantic search, and Git context.
keepygaga AGENTS.md
AGENTS.md instructions for TimWongUp/keepygaga, covering keepygaga agent entry, scope and authority, repo-native context, project constraints and commands and verification.
memory-forest AGENTS.md
AGENTS.md instructions for hyungchulc/memory-forest, covering repository contract, scope, invariants and change discipline.
keepygaga CLAUDE.md
Claude Code instructions for TimWongUp/keepygaga, a project described as: A warm, local-first keeper of memories for AI agents.
fieldtheory-cli CLAUDE.md
Instructions for afar1/fieldtheory-cli, covering claude.md, commands, architecture, key files and data flow.
siyuan AGENTS.md
AGENTS.md instructions for siyuan-note/siyuan, covering agents.md, 1. non-negotiable constraints, do not hand-edit, verification and prohibited operations and 2. project-specific rules.