Rumen Changelog

Changelog

All notable changes to Rumen will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

No unreleased changes.

[0.4.3] - 2026-04-25

Changed

• Confidence scores are now context-size-normalized. The two confidence: computeConfidence(rs) call sites in src/synthesize.ts (lines 228 and 620) are now wrapped with normalize(computeConfidence(rs), rs.related.length). Insights synthesized from small clusters (≤ 1, < 5, < 15 related memories) land at the lower ceilings the v0.4.2 confidence.ts curve specified — the raw computeConfidence output is unchanged, but the value that ends up on the Insight object now caps appropriately. Existing rumen_insights rows are unaffected; new ones land at the normalized scale.
• computeConfidence is now exported from src/synthesize.ts so the test suite can verify the raw function directly. The five computeConfidence: … tests now call computeConfidence(rs) instead of makePlaceholderInsight(rs).confidence and assert the unscaled values they were always meant to. Two new integration tests verify the wrapped-and-normalized confidence on the placeholder path.

Notes

• This is the deferred Sprint 26 T3 integration that landed in v0.4.2 with the pure function only. Sprint 27 closed the loop. npm test 58/58 green.

[0.4.2] - 2026-04-25

Added

• JSON parse hardening in src/synthesize.ts — Haiku-synthesized-insight responses now go through a three-pass tryParseInsight strategy (strict JSON.parse → fence/slice extraction → comma + literal-newline repair) before falling back to the per-object regex rescue. Drops the placeholder fallback rate from the 19% (31/166) observed on the 2026-04-19 production kickstart toward < 5% on common Haiku malformations: trailing prose after the JSON, markdown code fences, trailing commas, literal newlines inside string values. Truly malformed responses still cleanly fall through to placeholder. Helpers (tryParseInsight, sliceFirstJsonBlock, repairCommonJsonIssues) are exported and unit-tested in isolation.
• src/confidence.ts — pure normalize(rawScore, contextSize) function plus NORMALIZE_VERSION constant. Maps a raw 0..1 score onto a context-size-aware ceiling: 0.4 at size ≤ 1, 0.7 at < 5, 0.9 at < 15, full range at ≥ 15. Clamps NaN / out-of-range inputs. Currently exported only — integration into synthesize.ts is the Unreleased item above. Seven unit tests in tests/relate.test.ts cover the curve.

Changed

• Test count grew from 49 → 56 with the new fixtures.

[0.4.1] - 2026-04-16

Added

• Full test suite — 41 tests across extract, relate, synthesize, surface.
• Rumen install guide (docs/INSTALL.md) and kickstart script.
• README refresh covering v0.4 roadmap, hybrid Relate, and cost controls.
• Hybrid Relate documentation: embedding behaviour, failure modes, and the keyword-only fallback path.

[0.4.0] - 2026-04-16

Added

• Hybrid embeddings in Relate. relate.ts now generates OpenAI text-embedding-3-large embeddings per signal with per-signal error tolerance: timeouts / 4xx / 5xx responses fall back to keyword-only search rather than aborting the whole job.
• Self-healing migration. migrations/001_rumen_tables.sql gains ALTER TABLE ... ADD COLUMN IF NOT EXISTS blocks so the schema upgrades cleanly from v0.2 without a separate migration file.

Changed

• extract.ts, synthesize.ts, and types.ts updated to carry the new embedding vector through the pipeline.

[0.2.2] - 2026-04-14

Changed

• Renamed references from Mnemos → Mnestra. Final naming after Ingram was rejected (corporate sponsor conflict). Compatibility doc and SQL fixture renamed; the scoped @jhizzard/mnemos package is deprecated.
• SQL schema unchanged (memory_* tables stay the same).

[0.2.1] - 2026-04-14

Changed

• Mnemos branding pass through README, CHANGELOG, and the compatibility doc (docs/ENGRAM-COMPATIBILITY.md → docs/MNEMOS-COMPATIBILITY.md).

[0.2.0] - 2026-04-14

Added

• Synthesize phase (src/synthesize.ts) — replaces v0.1’s placeholder insight text with real Claude Haiku generation. Wired into runRumenJob between Relate and Surface.
  • Batching: up to 3 signals per model call, returned as a single JSON object for token efficiency.
  • Confidence scoring combines max similarity, cross-project bonus, and age spread of related memories.
  • Citations: the Haiku prompt asks for [#xxxxxxxx] short-ID citations inside the insight text and returns the matching full UUIDs in cited_ids[]. Only IDs that appear in the related-memory set survive into rumen_insights.source_memory_ids.
• LLM budget guardrails. RUMEN_MAX_LLM_CALLS_SOFT (default 100) logs a warning and falls back to the v0.1 placeholder template for the remaining signals; RUMEN_MAX_LLM_CALLS_HARD (default 500) aborts the job cleanly (rows already written stay, no corruption). Token counts are logged per call as [rumen-synthesize] tokens=<n>.
• Graceful degradation when ANTHROPIC_API_KEY is missing. Rumen logs [rumen-synthesize] no API key, falling back to placeholder and produces the same insight rows as v0.1 — so the loop remains testable without a Anthropic account.
• CI integration test. New integration-test job in .github/workflows/ci.yml spins up an ephemeral Postgres 16, applies tests/fixtures/mnestra-minimal.sql
  • migrations/001_rumen_tables.sql, and runs scripts/test-locally.ts end-to-end. Asserts at least one rumen_insights row is produced.
• Insight and SynthesizeContext types exported from src/types.ts for consumers that want to drive Synthesize independently of the full loop.

Changed

• surfaceInsights now accepts Insight[] rather than RelatedSignal[], so the placeholder and real-Haiku paths share a single write layer. External callers that were passing RelatedSignal[] directly must either move to runRumenJob (which still takes care of Relate → Synthesize → Surface) or wrap their signals with makePlaceholderInsight first.

Dependencies

• Added @anthropic-ai/sdk@^0.30.1.

[0.1.0] - 2026-04-11

Initial release. Extract + Relate + Surface only — no LLM synthesis, no question generation.

WARNING: Rumen v0.1 writes to a rumen_insights table. It does NOT modify or delete any existing memory rows. Run against a TEST instance for the first two weeks of use. Do NOT point at production memory stores until validated.

Added

• runRumenJob(db, options) entry point that runs the Extract, Relate, and Surface phases end-to-end.
• Extract phase (src/extract.ts): pulls session memories from the last 24–72 hours out of Mnestra’s memory_sessions + memory_items tables, filters out sessions with fewer than 3 events, and returns structured signals.
• Relate phase (src/relate.ts): for each signal, runs memory_hybrid_search across all historical memories and keeps top-5 candidates with similarity > 0.7.
• Surface phase (src/surface.ts): writes a non-destructive rumen_insights row per signal with source_memory_ids[] populated. v0.1 uses placeholder insight text; v0.2 will replace this with LLM synthesis.
• SQL migrations:
  • migrations/001_rumen_tables.sql — rumen_jobs, rumen_insights, rumen_questions with indexes.
  • migrations/002_pg_cron_schedule.sql — pg_cron schedule that calls the Edge Function every 15 minutes.
• Supabase Edge Function entry point at supabase/functions/rumen-tick/index.ts. Deno-compatible, reads DATABASE_URL via Deno.env.get.
• Local development script at scripts/test-locally.ts.
• Raw pg Pool factory at src/db.ts wired for Supabase Shared Pooler IPv4 URLs.
• [rumen-*] logging convention enforced across the codebase ([rumen], [rumen-extract], [rumen-relate], [rumen-surface], with [rumen-synthesize] and [rumen-question] reserved).
• Cost guardrails (hardcoded in v0.1): max 10 sessions per run via MAX_SESSIONS_PER_RUN, skip sessions with <3 events, skip sessions that already have a rumen_jobs row.
• Mnestra compatibility document at docs/MNESTRA-COMPATIBILITY.md.
• CI workflow that runs tsc --noEmit and a basic SQL syntax check.

Not included (by design)

• No LLM calls. Rumen v0.1 makes zero network calls to Anthropic, OpenAI, or any other model provider.
• No synthesis — insight text is placeholder in v0.1.
• No question generation.
• No self-tuning.