Skip to content

ROADMAP.md

Latest commit

 

History

History
93 lines (75 loc) · 4.64 KB

ROADMAP.md

File metadata and controls

93 lines (75 loc) · 4.64 KB
Raw

Roadmap

Last updated: 2026-07-27. Covers the twelve months to 2027-07.

What Cortex intends to do next, and — just as load-bearing — what it does not intend to do. This is direction, not a commitment: the project may miss any of it, and will say so here rather than quietly dropping an item. Dates are deliberately absent; the ordering is the promise.

Where the project is today

v4.17.1, released 2026-08-02. 52 standalone MCP tools (55 with the optional upstream integrations), 9 lifecycle hooks, 36 neuroscience-grounded mechanisms against a 97-reference bibliography, running on a local SQLite store by default or PostgreSQL + pgvector when configured. The OpenSSF Best Practices passing badge was earned 2026-07-27.

1. Verification depth over feature count

The next release series adds no new mechanism families. The work is making the existing surface provably correct rather than merely green:

  • Statement coverage to ≥80% (74.83% at 2026-07-27, measured in CI). The gap is concentrated in modules whose call sites were never wired: those are finished — the intended caller is built — rather than deleted, and each arrives with contract tests. A coverage floor is then enforced in CI so the number cannot silently fall back.
  • Maximal practical warning strictness. Ruff currently runs its default rule set and pyright runs basic behind a per-rule ratchet. Both tighten, rule by rule, with the ratchet floors only coming down.
  • Mutation testing widened past the single demonstrated module in pyproject.toml, so the test suite is judged by mutants killed rather than lines executed.

2. Supply-chain and project-health hardening

  • OpenSSF silver, then as far toward gold as one maintainer can go. Gold requires a second person for bus_factor, two_person_review and contributors_unassociated; everything else is reachable alone.
  • OpenSSF Scorecard findings closed, starting with the pinned-dependency set — every action is already pinned by SHA; the remaining findings are the container and workflow-permission ones.
  • Signed version tags to complement the existing Sigstore artifact attestations, so git tag -v verifies as well as gh attestation verify does today.

3. Ingestion beyond Claude Code sessions

The typed document seam shipped in v4.16.0 (ParsedDocument → normalizer → write path) exists to be extended, not to stay at two adapters:

  • Live Confluence connector over REST, reusing the same parser and provenance stamping the offline export adapter already uses — only the byte source and the provenance URL change.
  • Provenance-first ingestion everywhere: every ingested page keeps source path and content-hash version, so re-ingesting a revised document updates rather than duplicates.

4. The research line

Two papers are drafted in-repo (docs/arxiv-thermodynamic/, docs/arxiv-context-assembly/). The intent is to submit them, and to keep the benchmark harness (benchmarks/reproduce.sh) as the reproducible artifact behind every number the papers and the README quote. Benchmark floors stay gated on the isolated container run, never on a live store.

5. Raising the bus factor

The project has one maintainer and says so in GOVERNANCE.md. Adding a second maintainer with repository admin is wanted, is the single change that would most improve the project's resilience, and is not something a roadmap can schedule unilaterally — it depends on someone showing up and staying. Sustained reviewers are the path in; the door is the issue tracker.

What Cortex will not do

  • No hosted service, no cloud sync, no account. Cortex is local-first by construction; a server-side store would invalidate the guarantee in PRIVACY.md that memory content never leaves the machine.
  • No LLM in the retrieval loop. Retrieval stays deterministic and local — embeddings plus lexical signals plus a small cross-encoder — so a recall is reproducible and costs nothing per call.
  • No heavyweight model dependency. The 22 MB embedding model footprint is a hard constraint; a gigabyte-scale model breaks the runs-on-your-machine promise regardless of what it would buy in accuracy.
  • No re-absorption of the visualization stack. It lives in hypermnesia-mcp-viz and reads this store read-only; Cortex stays a memory engine.
  • No backward-compatibility shims. Format changes ship as one-shot migrations, per the standing rule in CLAUDE.md.
  • No telemetry by default. OTLP export stays opt-in behind an explicit environment variable, and never carries memory content.