Skip to content

ROADMAP.md

Latest commit

 

History

History
97 lines (77 loc) · 2.81 KB

ROADMAP.md

File metadata and controls

97 lines (77 loc) · 2.81 KB

Mneme Roadmap

Phase 0 — Project spine

  • Define project language and core promises
  • Establish v1 scope: personal memory first
  • Create initial repo docs, architecture notes, and tracked issues

Packaging / operator UX spine

Goal: make Mneme easier to run, inspect, and deploy.

Deliverables:

  • clearer split between protocol/, runtime/, and distribution/
  • one obvious operator-facing CLI surface
  • tiny sanitized sample workspace
  • install/migration guidance
  • cron/systemd deployment examples
  • strict failure propagation and schema-backed validation

Success criteria:

  • Mneme feels like a product, not a bag of scripts
  • operators can discover the right entrypoint quickly
  • deployment docs are boring and repeatable
  • runtime checks fail loudly instead of faking success

Phase 1 — Raw evidence ingestion

Goal: collect the things an agent actually did and saw.

Deliverables:

  • session/chat ingest
  • markdown/doc ingest
  • repo/file ingest
  • normalized evidence model
  • source identity and timestamps

Success criteria:

  • raw evidence can be captured without losing provenance
  • every evidence item has enough metadata to trace it back to origin

Phase 2 — Memory compilation

Goal: turn evidence into structured, usable memory.

Deliverables:

  • entity extraction (people, projects, systems, repos, tasks)
  • decision/timeline extraction
  • memory compiler output format
  • source-linked notes and summaries
  • confidence/state markers (observed, inferred, stale, contradicted)

Success criteria:

  • memory artifacts are more useful than raw logs
  • important claims can be traced back to source evidence

Phase 3 — Retrieval and recall

Goal: let agents recover what matters when they need it.

Deliverables:

  • semantic + lexical retrieval
  • recall API/tooling
  • answer assembly with citations
  • scoped recall by person/project/topic/time

Success criteria:

  • agents can answer prior-context questions with evidence
  • recall quality beats ad-hoc note grep and fragile session memory

Phase 4 — Memory hygiene

Goal: keep memory from rotting.

Deliverables:

  • duplicate detection
  • contradiction detection
  • stale fact surfacing
  • repair/review workflows
  • provenance-aware diffing of memory artifacts

Success criteria:

  • memory quality improves over time instead of degrading

Phase 5 — Promotion to shared memory

Goal: safely move verified knowledge from local memory into shared memory.

Deliverables:

  • promotion rules
  • review/audit trail
  • shared namespace model
  • ownership and conflict-resolution rules

Success criteria:

  • shared memory is useful without becoming a rumor mill

Candidate first issues

  1. Define the evidence model
  2. Build the first ingest pipeline
  3. Design the memory compiler output format
  4. Implement retrieval with citations
  5. Add contradiction/staleness detection
  6. Define promotion rules for shared memory