v0.3.4 — Coordinated native memory release

Query-independent core memory, bounded lifecycle ledgers, and a coordinated Python/native-plugin release source.

Core baseline and query recall

• OpenCode, pi, and omp refresh a DB-backed core baseline for each admitted user run, separately from query recall.
• Core and recall have separate budgets. Recall excludes rows already loaded into the baseline and abstains when no additional evidence remains.
• Remember, forget, and retraction become visible without restarting the session; repeated turns do not accumulate duplicate injected entries.
• LUMINARY_CORE_TOP_N=0 disables baseline loading.

Bounded ledger retention

• Lifecycle maintenance retains the newest rows per exact scope: 10,000 episodes, 50,000 audit events, and 50,000 evidence rows by default.
• Configure LUMINARY_EPISODE_MAX_ROWS, LUMINARY_EVENT_MAX_ROWS, and LUMINARY_EVIDENCE_MAX_ROWS; 0 disables pruning for that ledger. Older history beyond a cap is deleted during maintenance.

Release consistency and verification

• Python and the three native plugin packages use version 0.3.4 from the same tagged source. The tag workflow builds and validates the native tarballs once, then uses those artifacts for GitHub assets and npm publication.
• This forward release replaces the inconsistent npm 0.3.3 builds without overwriting published versions or moving the old tag.
• The inspector default-store regression test now isolates both HOME and Windows USERPROFILE from the operator's home.

Known limitations

• Native OpenCode Web GUI discovery and Hermes classic CLI/TUI visual inspection remain unsupported; the optional browser companion is not a native replacement.
• Local Node 22 dependency installation warns that @opentui/core@0.7.0 declares Node 26.4.0 or Bun 1.3.14. Build, OpenCode type checks, UI/plugin tests, and the real-model gate passed locally; this does not establish support for every host/runtime.

Installation and artifact verification

Python 0.3.4 is published on PyPI. npm latest is 0.3.4 for pi, OpenCode, and omp.

All three npm tarballs are byte-identical to the GitHub assets attached here, and every packaged dist resource matches its copy in the PyPI wheel. Registry digests were checked for the wheel, sdist, and npm tarballs. npm publication used GitHub Actions trusted publishing with signed provenance; no local npm login or passkey was needed for the release.

v0.3.3 — Interactive Memory Inspector

Interactive memory inspection for supported agent hosts, with explicit ownership, scoped access, and confirmed changes.

Native inspectors

• OpenCode CLI and Desktop, pi, and omp provide native inspection on their verified host versions.
• Hermes Desktop and its web dashboard provide separate native inspector integrations. The loopback browser companion is optional and must be launched explicitly.
• The inspector includes scoped browsing, history, sharing permissions, maintenance previews, transfer, settings, and an offline deterministic summary.

Reliability and security

• Inspector subprocesses resolve storage from the same trusted launcher environment as their native context. A conflicting launcher home is denied; config files cannot supply trusted inspector control channels.
• An unavailable selected config no longer falls back to another store, and ~-prefixed SQLite paths resolve against the trusted process home before opening.
• Malformed non-object responses from the companion's inspector subprocess return redacted unavailable or unknown_outcome results instead of dropping the request. Downloads retain their existing unavailable-or-revoked error.

Known limitations

• OpenCode Web loads the server plugin but does not discover third-party GUI contributions. The companion is not a native Web plugin.
• Hermes classic CLI and TUI do not provide the authenticated session and scope lifecycle required for a native visual inspector. Existing provider and command integrations remain available.
• A source-workspace npm audit --omit=dev reported 26 moderate dependency entries involving upstream glob and minimatch advisories through OpenCode/OpenTUI dependencies. Exploitability in installed hosts has not been assessed.
• Local installation on Node 24.19.0 warned that @opentui/core 0.7.0 declares Node 26.4.0 or Bun 1.3.14 as its minimum; the local build and native tests passed despite the warning.
• All three npm packages now list 0.3.3, but their later publication used develop builds whose entry points differ from this tagged release's tarballs. Matching versions do not imply identical provenance. For Python 0.3.3, use the plugin tarballs attached to this release rather than the npm 0.3.3 builds. A coordinated new version is required to restore a single release source; existing registry versions will not be overwritten.

Installation and verification

The release workflow gates the Python package and downloadable plugin tarballs on native-package checks, the official Hermes SDK, and strict recall with the real default embedding model. Check PyPI and the assets on this release for actual availability; a successful source build alone is not publication.

See the support matrix for tested hosts and unresolved targets.

v0.3.2 — Native Agent Plugins & OpenCode V2

Native memory packages for supported agent harnesses, with an OpenCode V2 adapter and evidence-aware memory operations.

Native agent plugins

• Luminary-maintained packages for OpenCode, pi, and omp, built from the same canonical adapters as the Python hook installer.
• Native memory tools and accurate stored, duplicate, skipped, error, and deletion results.
• pi and omp expose /luminary commands and native terminal interactions; compatible clients can consume their serialized results.

OpenCode V2

• Targets @opencode/cli 2.1.4 and @opencode/plugin 2.1.4 with Plugin.define, Promise domain APIs, and the plural plugins configuration.
• Automatic recall reads the admitted native user message; optional curation runs after successful execution.
• Session-owned workspace resolution prevents one shared plugin instance from reading another workspace.
• The older npm 0.3.1 adapter remains the V1 target; the V1 Ecosystem submission pins that exact version.

Memory ownership and accuracy

• Shared UTF-8 JSON bridge enforces user, workspace, and harness ownership.
• Explicit remember and ID-scoped forget use real storage outcomes, not inferred success.
• Automatic durable writes require opt-in LLM curation grounded in the user message; assistant-only claims are rejected.
• Telegram activity uses destination-specific cursors and evidence-bearing results rather than attributing global new IDs to the current turn.

Installation and downloads

Install the Python bridge from PyPI:

python -m pip install --upgrade luminary-memory==0.3.2

The release assets contain compiled 0.3.2 plugin tarballs for OpenCode, pi, and omp, plus the Hermes skill zip. All three npm packages are published at 0.3.2, with latest pointing to that version: luminary-memory-opencode, luminary-memory-pi, and luminary-memory-omp. Their registry SHA-512 checksums match the release assets.

Verification

• Python 3.11/3.12, PostgreSQL integration, and native-package CI passed before the source merge.
• The real default embedding-model release gate passed before PyPI publication.
• The published npm OpenCode 0.3.2 package was installed through the real V2 CLI, loaded as an active package plugin in server 2.1.4, and stored a fact through the published PyPI wheel.
• Real OpenCode V2 native sessions, Python, and SQLite exercised scoped memory operations, automatic recall, grounded curation, rejection of assistant-only claims, and cleanup on plugin removal. Model/curation replies used deterministic localhost fixtures; this is not a model-quality benchmark or a desktop/web visual test.

Full changelog: https://github.com/alertxsto/luminary-memory/blob/main/CHANGELOG.md

v0.3.1

v0.3.1 — measurement and recall-calibration release

Release gate on public strict recall, per-model semantic support floor (0.74
for the default model), date-invariant gold evaluation, relevance separated
from grounding, long-run soak harness, extended gold fixture, and develop
reconciled with main's documentation redesign and enricher retry fix.

v0.3.0 — Memory Accuracy & Session Continuity

Major release: memory is now scoped, evidenced, and auditable end to end.

Scope isolation

• New scope module centralizes tenant/scope SQL predicates used by every recall path (SQLite, pgvector, graph, keyword, temporal)
• Legacy/global rows stay visible to scoped callers through a compatibility mode

Schema and storage

• New tables: episodes, claims, claim_evidence, memory_evidence, memory_events
• Cross-process active dedup via scoped unique content-hash constraint
• SQLite schema now evolves existing databases, not just fresh installs

Claims and evidence

• Structured claims with grounded evidence quotes, canonical keys, and full lifecycle (active, superseded, retracted)
• Evidence fail-closed in every recall mode: fabricated quotes cannot become answer support

Hermes provider

• Every accepted turn enters a per-session episode ledger before curation
• When durable recall abstains, recent exact-session episodes are injected as bounded untrusted reference (active-task continuity)
• Serialized post-turn reconciliation can capture a new grounded fact or supersede/retract a scoped claim
• Scoped JSONL transparency events without prompt content or credentials

Enricher

• Transport switched from urllib to requests; urllib is blocked by Cloudflare (HTTP 1010) on OpenAI-compatible gateways
• Failures surface as enricher_failed events instead of being swallowed into llm_no_curated_summary

Lifecycle and maintenance

• Cleanup, prune, consolidate, and runner hardening
• Authority repair utility with dry-run and backup before apply
• Telegram activity hook requires the ok envelope and retries failures

Docs and website

• Guides refreshed for the new surfaces; website content and styles rebuilt

Full changelog: https://github.com/alertxsto/luminary-memory/blob/main/CHANGELOG.md

v0.2.16

v0.2.16: complete dashboard config coverage + config reference docs

v0.2.15

v0.2.15: adaptive memory + core memory integrity

Features:
- Adaptive importance on recall: frequently-recalled memories are
  re-estimated immediately after the access bump (same estimate_importance
  as lifecycle), so they climb into the next turn's persistent-context
  block; pinned rules (>= 0.9) are never downgraded.
- Rule-aware query expansion: when the knowledge graph yields no entity to
  expand a short query, up to two keywords from a durable rule on the same
  topic are appended before embedding. Original query tokens stay, so
  recall quality cannot regress (MRR 1.0 re-verified).

Fix:
- Content-level anti-duplication: a memory whose text is already in the
  core block was re-injected by persistent context or recall when it had a
  different id. Dedup now tracks content hashes alongside ids, so identical
  text appears exactly once per turn across core/persistent/recall.

Tests (373 passed, 3 skipped, ruff clean, hermes smoke OK):
- Core integrity: sourced only from DB core tag, non-core high-importance
  never leaks, content independent of _injected_ids, 3-way anti-dup once.
- Adaptive: recalled outranks idle, pinned never drops, defensive fallback
  paths (get_many/update_importances None, empty, no-change).
- Query expansion: rule keyword appended on topic overlap, unrelated query
  unchanged, short-word/empty-extra/error fallbacks.

Docs: README, docs/{index,recall,lifecycle,hermes-integration}, hermes/
SKILL+README, benchmarks/RESULTS, website, CHANGELOG, ROADMAP. Version
bump 0.2.15 in pyproject, __init__, plugin.yaml.

v0.2.13

feat: core memory (DB-backed MEMORY.md) + website redesign + test scr…

v0.2.12

feat: v0.2.12 — performance optimization pass (accuracy preserved)

- Vectorized rule auto-replace (scan_embeddings_matrix, ~19x faster, identical results)
- Lean persistent-context scan (top_by_importance, ~20x faster, no embedding decode)
- Batched access bookkeeping (touch_memories), lifecycle passes (delete_many/update_importances), temporal fetch (get_many)
- Docs/website/changelog/roadmap/benchmarks updated to latest data (93% coverage, 0.2.12, persistent context + rule hygiene)

v0.2.11

chore: release v0.2.11 — persistent context injection, rule pinning +…