memfleet-skills

MemFleet

Structural intent coordination for agent fleets.
Typed intent, precomputed blast radius, O(1) coordination reads — no prose, no timing windows.

Early Access — Phase 1 (intent registry, episode store, NodeState rollup, Class A/B/C, Y-doc, stdio MCP) is stable. Phase 2 (leases, shadow overlays, intent-aware auto-merge) and Phase 3 (in-process Memtrace, Ed25519 signing, CEL policy) are in flight. Report issues here.

---

The problem

Ten coding agents editing the same repo. The naive options:

• Prose rationale — each agent writes a paragraph. Others read it. 200+ tokens per edit, does not scale, vendor-specific, resists structural querying.
• Timing-based notification — broadcast in a 500 ms window, others subscribe. Works for two agents, floods at ten, breaks on network jitter.

Both produce the same outcome: silent collisions, overwrites, wasted tokens re-reading work that could have been predicted.

The solution

MemFleet treats intent as a structural type, not a paragraph. Every edit emits a typed IntentKind enum with a precomputed impact set, attached to the graph nodes it touches. Any agent reading those nodes receives the coordination picture — O(1), no prose, no timing window.

npm install -g memfleet     # binary + 10 skills + MCP server — one command
memfleet                    # MCP server auto-launched by Claude / Cursor on connect

Claude Code, Claude Desktop, and Cursor (v2.4+) pick up the skills and MCP tools automatically.

Typed intent — the core contribution

9 variants, 41 leaf values:

pub enum IntentKind {
    Refactor    { pattern: RefactorPattern },
    FeatureAdd  { surface: FeatureSurface },
    BugFix      { defect: DefectClass },
    Cleanup     { kind: CleanupKind },
    Performance { axis: PerfAxis },
    SecurityFix { severity: Severity, cve: Option<String> },
    TestAdd     { covers: Vec<NodeIdentity> },
    DocsOnly,
    Exploratory,
}

One intent ≈ 20 tokens. One prose rationale ≈ 200+. Across a 10-agent fleet running 100 edits — ~90,000 tokens saved per fleet-turn.

Conflict classes

Class	Meaning	Broker action
A	Additive only — new symbols, no removals	Auto-accepted. Zero coordination cost.
B	Modification to an existing symbol — body / type / rename	LWW by reference_time. Loser receives a typed ReplanHint.
C	Removes a live symbol or changes a signature with active callers	Blocked at intent time. Structured conflict report.

Refactor(RenameSymbol) + BugFix(NullHandling) on the same symbol are orthogonal — the broker auto-merges (Phase 2).

8 MCP Tools

Tool	Purpose
publish_intent	Register structural intent before an edit; returns blast radius and active conflicts
record_episode	Record a structural episode after an edit; classifies A/B/C, precomputes impact, updates rollups
get_node_state	O(1) read of recent episodes, active intents, dominant intent, conflict density for a symbol
get_episode	Fetch a single episode by id
query_episodes	Filtered episode search by node / intent / time range
ydoc_read	Read the Y-doc thread + NodeState blob for a symbol
subscribe	Register a streaming subscription with filters and budget
fleet_status	Active intents, open subscriptions, episode count, conflict counts by class

10 Agent Skills

	Skill	You say...
Intent	memfleet-publish-intent	"I'm about to refactor X"
Episode	memfleet-record-episode	"I just edited X"
Node state	memfleet-node-state	"is anyone working on X"
Query	memfleet-query-episodes	"what changed today"
Subscribe	memfleet-subscribe	"watch the auth module"
Status	memfleet-fleet-status	"how busy is the fleet"

Plus 4 workflow skills that chain multiple tools: