Design documents for AgentM, organized by parent/child with high-level or parent designs (HLD) followed by their children. This is a complete index of all current and proposed designs.

The Architecture page lists those designs that have been implemented but organized by architectural pillar for clarity.

Foundations — (Final) lays out the nine principles that guide the development of agentm and crickets, and how the person (agentm) and its tools (crickets) relate. Both HLDs inherit it by reference.

AgentM HLD — is the high-level parent design for AgentM. It is an overview of it's four pillars (Memory · Experience · Opinions · Personas) and how they are built on a durable memory engine, and where each touches crickets.

Design	What it covers	Status
Memory System	The engine end to end — how memory is stored, written, and recalled	Final
↳ Capture	Getting a thought or article into memory from your phone, browser, or a chat — the staging front door	Final
↳ Auto-Organization	Keeping the vault tidy on its own — write-time linking, tidying by age, dedup, and lint	Final
↳ Filing v2	The whole-vault convergence — layout, write authority, the calendar, memory lifecycle, and the dreaming binary	Final
↳ Vault	The vault shaped to the operator — the layout, the memory card, capture, dreaming, projects and tasks, a policy per space, one memory on every surface	Final
Memory↔Storage Seam	The contract memory uses to reach storage, so the backend can change without touching the engine	Final
↳ Vault Storage & Presentation	Where the vault lives and how it syncs to your devices	Proposed
Memory index	The metadata table behind fast, hybrid recall	Final
↳ Recall trace	Why a memory surfaced — the evidence behind each recall hit, kept after the session ends	Final
↳ Recall ledger retention	Expiring recall evidence on a clock, so what the ledger discloses stays bounded	Final
Experience & Dreaming	How AgentM learns between sessions — reflection, scheduled learning, dreaming	Final
Runner	The background-job executor that runs scheduled work on the host's own scheduler	Final
Goal contract	The contract for a persona pursuing an objective on an autonomous run	Final
Opinions & Gates	The standards AgentM holds, and asking for one by name	Final
Opinion registry	The resolver, the entry schema, and the catalog it serves	Final
Model + effort routing	Matching each role to the right model and reasoning effort	Final
Personas	The roster of roles, the gate that admits them, and how they launch	Final
Persona activation	Selecting and adopting a persona at runtime	Final
The persona tier	The classification the roster sits in — a standing concern that composes capabilities it doesn't own	Final

The original durable-memory design and its six implementation parts. Its still-live substance is being split into the memory designs above and then retired; kept here until that pass completes.

Design	What it covers
MemoryVault	The original design for durable, file-based agent memory
↳ write-primitives	How entries are captured and shaped
↳ recall-loop	How the right entries surface within a token budget
↳ reflection-and-recovery	Self-maintenance and crash recovery
↳ idea-ledger	The incubator for not-yet-acted-on ideas
↳ seed-pass	The co-created first pass that bootstraps a vault
↳ discovery-mining	Surfacing latent structure from existing notes

The three historical arc HLDs — Agent Memory Evolution, Device-Wide Architecture, and Memory-OS Architecture — were archived to the vault. Their still-live framing now lives in the AgentM HLD (the evolution-arc subsection) and Foundations (the device-wide substrate).

Architecture · Home