Memory protocol

Skill IrfanSadiqRahat/constellation/methodology/memory-protocol

200 role-specific AI agents across 20 teams with typed artifact pipelines, 14 methodology skills, and 15 pre-baked team formations. The virtual engineering org for Claude Code, Cursor, Codex CLI.

Install

npx -y skills add IrfanSadiqRahat/constellation --skill memory-protocol

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Persistent role-scoped memory across sessions. The agent that worked on your codebase yesterday knows things today.

SKILL.md

1.9 KB, 399 tokens by cl100k_base, as published. Nobody here has run it

memory-protocol

The problem

Stateless agents re-learn your codebase, your conventions, your past decisions every session. That's tax. Constellation's memory protocol lets each role accumulate role-scoped notes that survive sessions.

Storage layout

.constellation/memory/
  <project_id>/
    <role_id>/
      decisions.md       # ADR-style, append-only
      conventions.md     # codebase-specific patterns
      gotchas.md         # surprises learned the hard way
      cache.json         # structured fact cache (TTL'd)

Read rules

At session start, for role R:

Load .constellation/memory/<project>/<role>/* into the briefing
Cap to N lines per file (default 200) — older content stays on disk
Decisions and conventions are authoritative — never silently overwrite

Write rules

End of session OR on confirmed decision:

Append, don't rewrite. Decisions get superseded entries, not edits.
Sign each entry with timestamp + run-id
Compact opportunistically — merge near-duplicates with a compactor agent

Cross-role boundaries

Role memory is private to that role by default
Shared memory under <project>/shared/ for cross-role facts (e.g. tech stack, key URLs)
Privacy: PII never written to memory; redact on write

Forbidden

Hallucinated memory — only write what was observed or decided in this session
Memory as gossip — facts about teammates / users
Memory as PII store — actively redacted

Why this matters

Constellation's value compounds across sessions. After a month, the backend-architect "knows" your repo. After a year, the security-architect remembers every threat-model decision and its outcomes. Constellation gets more useful the longer you use it.

Gives 0 of the 12 instructions most memory context skills give in 399 tokens

Counted across 674 of the 847 authors here whose files we hold, read 2026-08-06

inform the user when setup is completein 21 of 674, across 6 files
confirm the draft with the user before writingin 21 of 674, across 6 files
update the agent skills block in place if it existsin 21 of 674, across 6 files
present findings to the userin 20 of 674, across 5 files
write the three docs files from seed templatesin 20 of 674, across 5 files
ask the user about each decision one at a timein 19 of 674, across 4 files
edit CLAUDE.md if it existsin 18 of 674, across 3 files
explore current repo statein 18 of 674, across 3 files
do not overwrite user edits to surrounding sectionsin 18 of 674, across 3 files
back up the original file before overwritingin 16 of 674, across 8 files
keep the memory index under 200 linesin 15 of 674
Provide actionable steps and verificationin 13 of 674, across 2 files

Said here and by no other author read

cap file content to N lines
never overwrite decisions and conventions
append new entries rather than rewriting
sign each entry with timestamp and run id
merge near duplicates opportunistically
keep role memory private by default

Grouped from the skills themselves: near-identical wordings counted once, and counted by distinct author, so one author publishing three of these counts once. Length counted with cl100k_base; the agent that loads this file may tokenize it differently.