Roadmap

Roadmap

What shipped in v3-alpha, what we’re still hardening, and the research bets behind it. Updated alongside every release.

TL;DR: Mnemosyne is v3-alpha — the “Digital Brain” expansion is implemented and wired at server boot (Apache-2.0, self-hostable). The 10 typed primitives, 26 MCP tools, gRPC, WebSocket, federation, consolidation, and plugins all ship today. Interfaces may still change before 3.0 stable.

Track legend

SymbolMeaning
✅Shipped — implemented and wired at boot in v3-alpha
🟡Hardening — shipped but interfaces may change before 3.0 stable
🔵Designed — spec exists, no code yet, open for feedback
⚪Research — exploring direction, no commitment

Foundations — what ships today

Everything below is implemented and exercisable against a running server.

Foundational primitives

CapabilityStatusRead
Bitemporal facts (valid_from / valid_to)✅Concept
Decisions, episodes, entities, relations✅Primitives
Multi-tenant RLS + FORCE✅Internals
Postgres + pgvector HNSW✅Recall pipeline

Recall & retrieval

CapabilityStatusRead
Hybrid recall (FTS + vector + entity drawer)✅Recall pipeline
Three-tier cascade (hot / warm / cold)✅Concept
Modes A / B / C (degrade gracefully)✅Concept

Governance

CapabilityStatusRead
Memory Worth (success/failure ratio)✅Concept
Consolidation Gate✅Concept
Zero-LLM features (anticipation · meta-cognition · composition · consensus)✅Concept

Interfaces

CapabilityStatusRead
REST API (~130 endpoints across 40 route modules, OpenAPI 3.1)✅API reference
MCP server (26 tools)✅MCP tools
gRPC (ConnectRPC, 17 RPCs)✅Spec
WebSocket subscriptions✅Spec
TypeScript SDK✅/sdks/typescript
Python SDK✅/sdks/python
Go SDK✅/sdks/go

v3-alpha — the Digital Brain (shipped)

The “Digital Brain” expansion is implemented and wired at server boot. Every section below has its own concept page. Interfaces are still labelled alpha and may change before 3.0 stable, but the code ships today.

Memory model expansion

ConceptStatusRead
Session memory + collapse-to-recall✅Read
Working memory (Redis ephemeral L0 layer)✅Read
CoT distillation engine✅Read
Self-editing memory (agent autonomy)✅Read
Procedural memory + prompt evolution✅Read
Hypergraph relations (n-ary associations)✅Read
Delta-encoded episodes✅Read
Tasks & events as primitives✅Read

Zero-LLM intelligence (the moat)

These four features turn an agent’s “let me think about that…” LLM tokens into deterministic SQL / code. Each one earned its own page.

ConceptStatusRead
Predictive memory (anticipation engine)✅Read
Meta-cognition (self-model + confidence map)✅Read
Memory compositionality (inference engine)✅Read
Trust-anchor consensus (zero-LLM conflict resolution)✅Read

Platform & distribution

ConceptStatusRead
Agent RBAC + namespace policies✅Read
Federation protocol (multi-instance sync)✅Read
Plugin architecture✅Read

Transports

Mnemosyne is no longer REST-only — it exposes a multi-transport surface, all shipped:

InterfaceStatusRead
REST (~130 endpoints)✅API reference
WebSocket subscriptions✅Spec
gRPC (ConnectRPC, 17 RPCs)✅Spec
MCP server (26 tools)✅MCP v3 tools

What’s next

v3-alpha lands the full Digital Brain surface. The work ahead is hardening toward 3.0 stable — locking interfaces, expanding test coverage, and acting on feedback from the alpha. The features above are wired and exercisable today; what changes before stable is polish and API stability, not whether they exist.


Want to influence what we build next?

  • Comment on a design: each concept page has an “Edit this page on GitHub →” link. Open a PR with proposed changes or open a Discussion.
  • Need something specific? Email contact@mnemosyne.dev — concrete use cases move things up the queue.
  • Want to sponsor the path to 3.0 stable? Email hi@mnemosyne.dev — sponsorship can change priority ordering.

The implementation is open (Apache-2.0). Pull requests welcome.