Provider Comparison

The attached report supports a compositional design rather than adopting one provider wholesale. This page records donor patterns only; it does not add these systems as dependencies.

Donor Patterns

Source pattern	Useful component	Roadmap role
Hermes built-in memory	MEMORY.md and USER.md remain active	Current V1 L0 core memory alongside the external provider.
Letta	Always-visible memory blocks	Current V1 discipline for pinned core memory.
Memory-OS	Trust and injection policy	Current V1/V1.1 discipline for cited, bounded injection and verification warnings.
Memori	Attributed background capture	Current V1 evidence-ledger and event-attribution pattern.
RetainDB	Scoped context and agent events	Current V1 context-routing pattern; V2 handoff pattern.
Mem0	Typed fact modeling and retrieval patterns	Current V1 semantic fact plane design.
Supermemory	Profile and hybrid memory/RAG surfaces	Future profile-serving pattern.
TencentDB-Agent-Memory	Compact symbolic session state	Current V1.1 active task-state projection pattern.
OpenViking	URI-like browseable memory/resource model	Current V1.1 workspace resource tree pattern.
ByteRover	Local, versionable context tree	Current V1.1 inspectable project-memory pattern.
Graphiti/Zep	Temporal graph with provenance	V2 relation graph and validity-window pattern.
Hindsight	Reflection over stored evidence	V2 slow-path synthesis pattern.
Honcho	Peer and identity reasoning	V3 social-memory sidecar pattern.

V1 Decision

V1 keeps the implementation simple:

• one Hermes provider;
• TypeScript-owned memory behavior and persistence;
• Python only at the Hermes adapter boundary;
• SQLite + FTS;
• append-only evidence;
• typed facts;
• active session state;
• workspace resources;
• context packs;
• verification records;
• no graph, vector, cloud, or LLM extraction dependency.

The provider landscape should be rechecked against primary docs before any donor pattern becomes an implementation requirement.

Future Mapping

• Later local additions must stay local-first and dependency-light.
• V2 may add temporal graph, contradiction handling, reflection, and handoff packets when recall safety requires them.
• V3 may add social memory, connector sync, shared memory blocks, and federation after the local system is proven.

Anti-Pattern To Avoid

Do not collapse evidence, summaries, facts, profiles, and beliefs into one store. Evidence should be append-only. Facts should be cited and updateable. Context packs should be generated views. Temporal and reflective reasoning belongs in later projections.