Skip to content

ROADMAP.md

Latest commit

 

History

History
227 lines (168 loc) · 5.34 KB

ROADMAP.md

File metadata and controls

227 lines (168 loc) · 5.34 KB
Raw

TurboMemory Roadmap 🚀

Claude-style long-term memory with 4/6/8-bit TurboQuant compression — runs on a laptop.

Vision

TurboMemory aims to become:

  • SQLite-like simplicity for semantic memory
  • Parquet-like storage efficiency for embeddings via TurboQuant
  • Edge-first replication for distributed memory networks
  • Self-healing memory via background consolidation

Current Status: v0.5 (✅ Released)

  • Core memory engine with SQLite + quantization
  • Packed 4-bit / 6-bit / 8-bit embedding storage
  • Topic-based segmentation with centroid prefilter
  • Retrieval verification with cross-checking
  • Quality scoring + decay
  • Exclusion rules (what NOT to store)
  • Self-healing consolidation (merge/prune, contradiction resolution)
  • Observability + metrics
  • Plugin system for extensibility
  • LangChain integrations (retriever, chat history)
  • CLI with stats, search, verification flags
  • Streamlit dashboard scaffold
  • TMF v1 storage format specification
  • Hybrid search (BM25 + vector fusion)
  • FastAPI server with multi-tenant support

v0.5 Release includes: TMF v1 format, verify CLI, Parquet/Lance export, hybrid search, server mode

v0.5 — Storage Format Specification (TMF v1) ✅

Goal: define a stable storage format (Parquet-like for semantic memory).

Deliverables

  • Define "TurboMemory Format (TMF)" v1 spec:
    • .tmindex (SQLite metadata index)
    • .tmvec (packed vectors file)
    • .tmlog (append-only transcript/event log)
    • .tmmeta.json (schema + model metadata)
  • Schema versioning + migrations
  • Checksum verification + corruption detection
  • Deterministic serialization format for vectors
  • Fast export/import tooling
  • CLI tm verify command

Metrics

  • Storage can be copied between machines and loaded instantly

v0.6 — Hybrid Search (Vector + Keyword) ✅

Goal: become usable for real RAG and enterprise search.

Deliverables

  • BM25 or keyword search fallback
  • Scoring fusion:
    • vector similarity score
    • keyword score
    • recency score
    • confidence score
  • Metadata filters:
    • time range
    • topic filter
    • tags/namespace filter
  • Query explain output (debug mode)

Metrics

  • Improved retrieval accuracy on real datasets
  • Stable query API supporting filters

v0.7 — Server Mode (Microservice API) ✅

Goal: allow TurboMemory to run as a service.

Deliverables

  • REST API (FastAPI):
    • /add
    • /bulk_add
    • /query
    • /delete
    • /stats
  • Multi-tenant namespaces
  • API key auth
  • Docker image
  • Rate limiting / request validation

Metrics

  • Stable service running on VPS with <500MB RAM

v0.8 — Replication & Sync (Edge-first)

Goal: local-first semantic DB that syncs like Git/WAL.

Deliverables

  • Log-based replication:
    • Node A asks Node B for missing offsets
    • Node B streams missing .tmlog events
  • Conflict handling policy:
    • Append-only ID uniqueness
    • Merge rules for duplicates
  • Sync over HTTP
  • Optional encryption for replication payload

Metrics

  • 2 nodes can sync 100k memories reliably
  • Idempotent sync (safe to retry)

v0.9 — Index Acceleration (Optional HNSW/IVF)

Goal: make retrieval fast for millions of vectors.

Deliverables

  • Optional HNSW index module
  • IVF centroid bucket acceleration
  • Caching layer for hot topics
  • Multi-thread query execution
  • Vector block prefetching

Metrics

  • Query latency remains low at 1M+ chunks on CPU

v1.0 — Production Core Release

Goal: stable foundation usable in real apps.

Deliverables

  • Stable API freeze
  • Full documentation site
  • Full test coverage for packing/indexing
  • Verified storage integrity guarantees
  • Stable migrations
  • Release notes + changelog discipline

Metrics

  • Safe upgrades
  • Reproducible performance

v1.1+ — Distributed Sharding / Cluster Mode

Goal: scale horizontally across machines.

Deliverables

  • Sharding by:
    • topic
    • time range
    • centroid hash
  • Router node (fan-out query)
  • Distributed top-k merge
  • Node health metrics + monitoring endpoints

v2.0 — Semantic Data Lake Index

Goal: semantic indexing layer for S3/object storage.

Deliverables

  • S3-compatible backend for .tmvec and .tmlog
  • Scalable metadata store option (Postgres)
  • Caching proxy nodes
  • Batch ingestion pipelines
  • Integration examples with LlamaIndex/LangChain

Contributor-Friendly Areas

If you want to contribute, these are high-impact modules:

Beginner (Good First Issues)

  • Docs improvements
  • Tests
  • CLI UX
  • Examples and demos

Intermediate

  • Benchmark harness
  • Scoring fusion logic
  • SQLite optimization

Advanced

  • Packed embedding codec improvements
  • Replication protocol
  • HNSW/IVF index integration
  • Corruption recovery tools

Guiding Principles

TurboMemory is built around:

  • Local-first
  • Small footprint
  • Append-only logs
  • Cheap compressed storage
  • Self-healing consolidation
  • Portable file format
  • No cloud dependency

Long-Term Goal

TurboMemory should become:

"The default open storage format for semantic memory and compressed embeddings."

If you want to help build that future, join the project 🚀