V1, V2, and V3 Roadmap

This roadmap preserves the local-first V1 boundary while keeping the report’s larger Meta Memory OS direction visible. The product shape is one Hermes provider with many internal planes, not several competing providers.

stateDiagram-v2
  direction LR
  state "Phase 0: adapter contract alignment (local smoke)" as P0
  state "Phase 1: local-first V1 implementation" as V1
  state "Phase 1.1: local-first product hardening" as V11
  state "Phase 2: V2 Core safer recall" as V2
  state "Phase 3: shared and federated memory (design only)" as V3

  [*] --> P0
  P0 --> V1: local contract proof
  V1 --> V11: evidence, facts, FTS, context packs
  V11 --> V2: local recall-safety warnings
  V2 --> V3: after V2 is proven

Phase 0: Adapter Contract Alignment

Before adding runtime features, align the Hermes adapter boundary to the current local plugin contract and prove it locally:

• plugin.yaml metadata and required fields;
• root plugin shim for Git-based Hermes plugin installs;
• Python register(ctx) entrypoint and discovery shape;
• initialize(...) bridge setup;
• lifecycle hook names and current pre_tool_call, post_tool_call, and on_session_end registration;
• tool schemas and handler registration for status, context_pack, remember, search, V1.1 session/resource tools, and verify;
• optional setup-skill registration when Hermes exposes register_skill(...);
• local smoke proof for adapter compilation, manifest shape, initialize(...), register(ctx), tool schemas, status diagnostics, CLI delegation, configured SQLite path handling, and tool-call argument boundaries.

Until this is loaded and exercised by a real target Hermes version, adapter work remains a local contract alignment and smoke-test target, not a production Hermes installation claim.

Phase 1: Local-First V1 Implementation

V1 is the minimum auditable Hermes meta-provider:

• TypeScript domain model.
• SQLite schema and FTS indexes.
• Append-only evidence events.
• Core memory blocks.
• Semantic facts with source event IDs.
• Context packs with token estimates and citations.
• Verification records.
• CLI bridge for JSON stdin/stdout.
• TypeScript/Vitest tests for core, SQLite, and CLI behavior.
• pnpm and Turborepo orchestration.
• Thin Hermes Python adapter that maps Hermes hooks and tools to the TypeScript CLI only.

V1 does not include vector search, graph databases, cloud providers, LLM extraction, reflection, connector sync, social memory, or a second Hermes provider.

Phase 1.1: V1 Product Hardening

These are implemented as local-first, dependency-light projections:

• active session-state projection for compact current-task memory;
• browseable workspace/resource tree;
• richer context-router policy with auto, task, and workspace modes;
• local adapter smoke fixture for manifest shape, initialize(...), register(ctx), tool schemas, handler delegation, hook signatures, and observer tool-call evidence writes.

Any later local-first additions must stay auditable and dependency-light. Session state and workspace tree work must continue to avoid vector DBs, graph DBs, cloud memory providers, connector sync, and LLM extraction dependencies.

This is still not a production Hermes installation claim until a target Hermes checkout loads the plugin and exercises the hooks/tools end to end.

Phase 2: V2 Core Safer Recall

V2 Core is implemented as local-first recall-safety warnings:

• scoped temporal relation storage;
• active contradicts and supersedes warnings in context packs;
• source citation validation for packed facts, session state, and resources;
• optional current-branch, commit, missing-file, and file-hash drift warnings when context-pack requests include local workspace provenance;
• add-relation and probe-relations CLI commands plus matching Hermes tool delegation.

V2 Core does not filter risky memory. It keeps packed items visible and returns structured recallWarnings so the agent can inspect the risk.

Future V2 work remains design only:

• richer temporal graph reasoning and validity-window conflict handling;
• optional reflection for synthesis-heavy prompts;
• handoff packets if multi-agent workflows need them.

Future graph and reflection work should stay routed. It should run only when the query needs temporal, relational, or inferential reasoning.

Phase 3: Shared And Federated Memory

Social and peer memory, external connector sync, shared memory blocks, and federation hooks belong after V1 is stable and V2 recall safety is proven.

Operating Guardrails

• Evidence stays append-only.
• Facts stay typed, cited, and updateable.
• Context packs stay bounded and inspectable.
• Project and workspace memory stay separate from personal profile memory.
• Provider-landscape claims from the report remain roadmap guidance until they are rechecked against primary docs.