Compartment

• CollectionMemory & State
• Classificationagent-native
• LicenseApache-2.0
• MCPAvailable
• Agent SkillAvailable
• InterfacesAgent Skill · MCP · CLI · STDIO
• Verified2026-10-19

• Website ↗
• Docs ↗
• Official repo ↗
• Full dossier on GitHub ↗

What it does

Compartment is an encrypted, fully offline memory vault for AI agents. Official docs: one claim per memory (hard 200-character / no-list rule), required source + discovered date, optional expires, opinions that supersede rather than duplicate, hybrid vector+keyword recall (~12 ms) over an in-memory index, and a hash-chained audit log. Embeddings are a bundled ONNX model; no LLM inside, CI-enforced no network. Vectors are encrypted at rest; only the passphrase opens the vault.

A new vault includes ~6,700 reference facts (toggle-able). Capture does not depend on the model calling a tool: the Claude hook writes memory files even if the host’s system prompt overrides tool instructions.

Distinct from Claude-Mem / agentmemory: Compartment’s differentiator is offline + encryption + one-claim discipline, not session compression or a shared unencrypted server.

Agent-facing interfaces

• Interaction pattern: CLI + MCP (compartment serve)
• Catalog protocols: Agent Skill, MCP, CLI, STDIO
• MCP: ✅ Available — this is the agent surface

• Agent Skill: ✅ /compartmentalize installed by integrate claude

  hermes

  openclaw (not npx skills add)

• Onboarding type: mcp

How agents onboard

Interaction pattern: CLI + MCP (compartment serve)

pip install compartment && compartment init
compartment integrate claude      # also: hermes, openclaw, or --all
# MCP clients without a recipe:
# { "mcpServers": { "compartment": { "command": "compartment", "args": ["serve"] } } }

See the full onboarding section in the dossier.

Protocol surface

Verification

Last verified 2026-10-19.

Sources:

• https://api.github.com/repos/MaxFreedomPollard/Compartment