-
Notifications
You must be signed in to change notification settings - Fork 0
Designs
github-actions[bot] edited this page Jun 27, 2026
·
14 revisions
The high-level design docs behind Agent M — its memory layer and the harness tooling around it — the full HLDs for what shipped and what's in flight, where the design started and where it's going. The Architecture pillar overviews link down into these for the deep detail.
The permanent agent-memory store and its six implementation parts.
-
MemoryVault — the 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 + 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 current top-level design parents (AG-track Phase 1, status: launched). Children point up at these; all children are content-final and launched as of AG Phase 3 (2026-06-24).
- Foundations — the shared cross-repo "why": the nine principles agentm and crickets both stand on, and how the person (agentm) and its tools (crickets) relate. The root both HLDs inherit by reference.
-
AgentM HLD — the agentm parent: the four pillars (Experience · Memory · Opinions · Personas), their components, and where each touches crickets. Succeeds the V5 Memory-OS HLD (vault-archived 2026-06-24, AG Wave 2; its V1→V8 arc lives in the AgentM HLD's Evolution-arc subsection).
- Memory System — the seam, backends, write protocol, recall loop, storage layers, V5-14 target gap.
-
Memory↔Storage Seam — the storage contract (seam verbs,
Locatorguards, tiers), backends, V5-0 write protocol, routing plane, backend-aware harness state. - Experience & Dreaming — reflection (backward), scheduled learning (forward), heat policy, dreaming, the scheduler.
- Opinions & Gates — the nine-opinion catalog and the request-by-name model.
- Opinion registry — the request-by-name resolver, the entry schema, the catalog the resolver serves.
- Personas — the persona tier, the ~11-persona roster, the gate, launch modes, role-retirement.
-
Model + effort routing — the T0…T4 tier scale (Claude + Gemini), the persona→tier map, the
tier:manifest axis. - Runner — agentm's background-job executor on the hosts' built-in schedulers; writes routed by ownership tier, reported in the digest.
- The three historical arc HLDs — Agent Memory Evolution (V1→V8), Device-Wide Architecture (V4), and Memory-OS Architecture (V5) — were vault-archived 2026-06-24 (AG Wave 2); their still-live framing now lives in the AgentM HLD (the V1→V8 Evolution-arc subsection + the V5 baseline) and Foundations (the V4 device-wide substrate). Full text:
<vault>/_vault-archive/ag-design-history/. - The persona tier — the third classification above the substrate/plugin binary: a standing concern that composes capabilities it does not own (brain = degenerate persona; Planner (TPM) = first real persona). Refines the V5 unbundling (agentm-hld); folds the former ADR 0016 (its load-bearing calls).