Roadmap

Midas is eval-first: the roadmap is shaped by what the measurements say, not by a feature wishlist. This is a living document — open a Discussion or an issue to push on any of it.

The contract (from 1.0.0)

The public surface — the midas.Memory SDK (remember / recall / consolidate / supersession / forget), the guard (guard_reliance, decide_memory_use, the coding is_forbidden gate), the MCP tools, and the midas CLI — follows semver. Breaking changes only land in a major. Additive, no-LLM, measured improvements are always welcome within a minor.

Design invariants (these will not change)

These are the bet, not a backlog. They are load-bearing for the cost/privacy/auditability story:

• No LLM at ingest or query. $0 per message, zero data egress, deterministic. New no-LLM mechanisms are very welcome; an LLM in the ingest/query path is out of scope by design.
• Local-first. Memory is a SQLite file on your machine. No account, no telemetry, works offline.
• Source-traceable recall. Recall returns the verbatim stored turn, never an LLM rewrite.
• Governed, not just retrieved. Provenance + currency gate what an agent may act on.

Shipped (1.0.0)

The retrieval core (LongMemEval-s recall@k 0.92, LoCoMo 0.73, BEAM to 10M tokens), typed belief revision, selective forgetting, the provenance guard + memory-safety bench (ASR 0.00 / benign-pass 1.00), the coding-agent vocabulary, per-project scoping, the MCP server + midas CLI (init / serve / status / doctor / inspect / export / import), the machine-readable wiring receipt, and the TypeScript port. Full detail in CHANGELOG.md and BENCHMARKS.md.

Near-term (help wanted)

Ordered by leverage, not certainty. Several are good first issue-sized.

• Third-party head-to-heads. Run Mem0 / Letta OSS through eval/ on LongMemEval-s and LoCoMo at the same embedder budget and seeds, and publish the table with the cost column — including any row where Midas loses. This is the single highest-credibility measurement left to make.
• Judged answer-rate at scale. The current gpt-4o judged number is n=40 (cost-bound and honestly flagged). A cheaper judge at n≥200 × 3 seeds turns it from anecdote into a defensible number.
• More client guides. One page per client (Cursor, Claude Code, Codex, Windsurf, Zed, …) covering wiring, shared-memory setup, and verifying with midas doctor.
• Dataset mirrors & repro polish. Make every BENCHMARKS.md number reproducible from a clean checkout with a single documented command and a pinned dataset.
• Footprint at scale. Extend the measured RAM/disk figures to 100k / 1M records.

Explicitly not planned

Kept here so nobody has to guess — each is a documented, measured ceiling, not an oversight:

• LLM-summarized / abstractive memory at ingest. Measured to hurt (naive distillation 0.37 → 0.08) and it breaks the $0 / zero-egress / traceable invariants. See docs/frontier-2026.md.
• Whole-conversation aggregation / summarization queries. Top-k retrieval structurally can't cover these; documented as the trade Midas makes. See BENCHMARKS.md.
• A cloud service, account system, or telemetry. Midas is local-first by design: the memory is a SQLite file on your machine and it never phones home. Everything in this repo is and stays free under Apache-2.0.