Documentation14 entries
Everything about how Midas works — and why.
SDK reference, benchmark methodology, and the research notes behind governed memory. Every page mirrors a file in the Midas repository.
⌘K
Start
05 · entriesBenchmarks
02 · entries01
Benchmarks
Reader-independent retrieval and cost numbers, reproducible from the repo.
- TL;DR
- 1. Retrieval quality — recall@k + precision@k (deterministic)
- 2. Cost / latency — the no-LLM edge (memory layer only)
- 3. Provenance (auditability)
- 4. Scaling — sub-linear search past the exact scan (ANN)
- 5. Retention — selective forgetting beats recency (no LLM)
- 5b. Adversarial near-duplicates & temporal conflicts — conflicts-v1
- 6. Correctness with a fixed strong reader (secondary)
- 7. Governance & safety — deterministic, no-LLM (the trust axis)
- Methodology — why reader-independent metrics
02
Methodology
Why retrieval + cost lead, why the answer score is a secondary signal.
- 1. What the pipeline does NOT do
- 2. Disentangling retrieval quality from answer quality
- 3. Adversarial near-duplicates and temporal conflicts (conflicts-v1)
- 4. How conflicting memories are handled today
- 5. Selective forgetting: where it worked, where it failed
- 6. The exact memory policy injected over MCP
- 7. Reproducing everything on this page
Research
03 · entries02
Frontier 2026
Where agent memory is heading and Midas' position in it.
- 1. The strategic shift: the 10M-token regime
- 2. The frontier systems, and what Midas takes from each
- 2b. The distillation pattern — LIGHT's win without the LLM bill (measured-motivated)
- 3. What Midas deliberately does NOT adopt
- 4. Measured so far (deterministic, reproducible)
- 5. Roadmap (eval-first, in order)
03
Research notes
Working notes on retrieval, conflict handling, and forgetting.
- 1. End-to-end correctness is reader-dominated, not memory-dominated
- 2. Hosted MoE judges are not reproducible across sessions
- 3. A weak local reader cannot measure memory quality
- 4. "No LLM at ingest" is a measurable cost/privacy wedge
- 5. Reranking is not always on the retrieval-quality path
- 6. The retrieval ceiling is the embedder, not the budget
- 7. Hybrid lexical+semantic didn't beat the bi-encoder here (a clean negative)
- 8. Where the default store tops out (~10K memories)
- 9. Past the exact scan: a numpy-only IVF index, and the crossover that says when to use it
- 10. Dumb reader + adversarial conflicts — isolating retrieval from the reader