Why this exists
Agents are useful, but they miss the context that makes work correct.
They may know the task and still miss the person, project, relationship,
decision, or format behind it.
Marshmallow gives them source-backed recall: the few entities, decisions,
working rules, and open loops that matter now.
It keeps a small, source-backed graph of the things you explicitly ask it to
learn, then routes that graph into compact indexes, task-shaped recall packets,
runtime guidance, and optional skill overlays.
Marshy is the mascot. Marshmallow is the system.
Cute face, boringly inspectable files underneath.
The idea
sources -> typed graph nodes -> indexes/recall packets -> runtime.md -> adapter -> agent
- Sources are things you chose: files, notes, examples, rejected outputs,
corrections, screenshots, PDFs, or URLs.
- Graph nodes are compact, source-backed records: entities, decisions,
relationships, preferences, and working rules.
- Indexes are agent-written navigation pages that keep future agents from
crawling the whole graph.
- Projections are task-shaped recall packets for meetings, handoffs,
workflows, or focused agent work.
runtime.md tells the agent to check indexes first, then load only the
graph nodes and projections that matter now.- Adapters connect that runtime file to
CLAUDE.md or AGENTS.md.
- Skill overlays are optional downstream tuning for existing agent skills.
The goal is not to make a giant memory layer. The goal is to give connected
agents the right source-backed context before they draft, decide, or act.
Trust model
Marshmallow is deliberately boring where trust matters.
- Local-first. It writes plain files under
~/.marshmallow/.
- Explicit learning. No background capture. No silent session ingestion.
- Source-backed guidance. Graph nodes point back to real sources or approved
corrections.
- Preview before mutation. Adapter installs and skill rewrites show you what
will change.
- Rollback. Applied mutations create byte-exact backups and rollback records.
No hosted profile. No dashboard. No database. No memory daemon.
Quickstart
Claude Code
/plugin marketplace add notmehul/marshmallow
/plugin install marshmallow
Prefer the CLI? claude plugin marketplace add notmehul/marshmallow && claude plugin install marshmallow
Then start the calibration:
/marshmallow:start
Marshy asks for a small context pack: people, projects, decisions, formats,
corrections, rejected outputs, or working rules. Marshmallow turns that bundle
into the first source-backed graph, previews the runtime adapter, and can propose
optional skill overlays when they are useful.
Nothing durable is written without your explicit approval.
Later, teach it more with /marshmallow:learn, find context with recall, and
retune skills with /marshmallow:tune when a reusable skill should change.
Codex & Cursor
Marshmallow's graph and recall packets are plain files. Codex and Cursor can
read the same context through an AGENTS.md adapter:
scripts/marshmallow.py init
scripts/marshmallow.py adapter preview --harness codex # ~/.codex/AGENTS.md
scripts/marshmallow.py adapter apply --harness codex
scripts/marshmallow.py adapter preview --harness cursor # ./AGENTS.md
scripts/marshmallow.py adapter apply --harness cursor
What it creates
~/.marshmallow/ is the source of truth — plain files, no database:
runtime.md # short instructions imported by CLAUDE.md / AGENTS.md
inbox/ # unsynthesized candidate material (untrusted until promoted)
sources/ # source cards with pointers and provenance
graph/ # source-backed context nodes (the durable substrate)
indexes/ # compact navigation pages for agents
projections/ # task-shaped recall packets
overlays/ # approved skill alignment overlays
backups/ # exact backup bytes plus record.json for rollback
Skills
