memorix 1.4.5 → 1.5.0

package/CHANGELOG.md

@@ -2,6 +2,84 @@
 
 All notable changes to this project will be documented in this file.
 
+## [1.5.0] - 2026-08-14
+
+### Added
+- **Stability stress suite** -- Five deterministic, offline stress exams now
+  run in the normal test suite: a 2,000-observation corpus (planted needles,
+  ranked determinism, project isolation), 60-way parallel writes plus
+  20-way same-topic upserts, session churn across projects with parallel
+  starts, 300-record long-term maintenance exactness, and repeated CLI
+  invocations. Live embedding stress exams are env-gated
+  (`MEMORIX_RUN_LIVE_EMBEDDING_TESTS=1`) and never run in CI.
+- **Provider-aware embedding default** -- When the embedding base URL
+  points at OpenRouter and no model is configured, the default is now
+  `qwen/qwen3-embedding-8b` (4096 dimensions) instead of the OpenAI-only
+  default. An explicitly configured model always wins.
+
+### Changed
+- **Verified OpenRouter embedding lane** -- The `api` embedding provider
+  was exercised live against `qwen/qwen3-embedding-8b` on OpenRouter:
+  a 100-text batch, cache round-trip, and provider-singleton stability all
+  pass. Docs cover the env-var form and borrowing `OPENROUTER_API_KEY`.
+- **Multimodal vision option documented** -- Image analysis can run on the
+  LLM lane through an OpenAI-compatible vision endpoint, including
+  OpenRouter models such as `qwen/qwen3-vl-8b-instruct`.
+
+## [1.4.6] - 2026-08-14
+
+### Added
+- **Memory-native brief block** -- Every task brief opens with a "You and
+  this workspace" block: the user profile (personal `user-profile`
+  observations), the latest ended session summary, and recent durable
+  long-term memories, all inside the existing token budget. Older callers
+  and fixtures without the block are unaffected.
+- **Long-term memory self-maintenance (rule leg, no LLM required)** --
+  Explicit long-term requests (MCP `longTerm: true`, CLI add/promote)
+  auto-qualify on the spot instead of waiting for a manual review;
+  hook-captured and Git-derived candidates keep the review path. Stale
+  qualified/approved records auto-archive after 30 days without activity,
+  and a newer record auto-supersedes same-title records in the same scope
+  and kind. Maintenance runs in the existing durable lane at session start,
+  after each explicit promotion, then on a daily heartbeat. Approval stays
+  an explicit operator action.
+- **`memorix purge`** -- Retires the current project's memories from
+  retrieval (`--all` for every project, `--yes` required outside a
+  terminal). Observations and long-term records are archived, never
+  hard-deleted, so the audit trail survives.
+- **Eval coverage** -- New deterministic, offline exams pin the always-on
+  brief, long-term maintenance, profile parity, purge safety, git ingest
+  policy, media schema honesty, and server config defaults.
+
+### Changed
+- **Installed tool profile matches taught tools** -- `memorix setup` and
+  every plugin template now run `memorix serve --mode lite`, so the tools
+  taught in generated guidance (sessions, timeline, reasoning, retention)
+  actually exist. Repo guidance marks the full-only tools
+  (`promote`, `rules_sync`, `workspace_sync`) with `MEMORIX_MODE=full`.
+- **Honest media schema** -- With the generation/transcription gates
+  closed, `memorix_media` no longer advertises `generate-image`,
+  `generate-video`, or `derive-audio` (or their parameters); agents see
+  only what the operator enabled.
+- **Config that does something** -- `[server].port` and
+  `[server].dashboardPort` are now real startup defaults for `serve-http`
+  and `dashboard`; `git.ingest_on_commit` actually gates post-commit
+  ingestion and `git.max_diff_size` caps captured diff content. The
+  never-consumed `mcpServers` key was removed from `memorix.yml`
+  (unknown keys are ignored), and `[team]`/`[hooks]` are documented as
+  parsed-but-reserved.
+
+### Fixed
+- **Guidance hygiene** -- Generated agent guidance now teaches
+  store-the-underivable, dual-channel feedback, verify-before-use, the
+  ignore contract, user-profile capture, and session-end summaries.
+- **Session search dedup** -- `memorix_search` demotes rows already shown
+  in the same session (duplicate rate 25% → 19% in the replay exam, hits
+  and token budget unchanged).
+- **Dead code removal** -- Removed the superseded slash-REPL, the legacy
+  in-memory team layer, retired TUI components, and duplicate commands
+  (about 4,600 net lines), each verified unreferenced before deletion.
+
 ## [1.4.5] - 2026-08-14
 
 ### Added

package/CLAUDE.md

@@ -61,6 +61,10 @@ Use `memorix_resolve` when a task is done or a bug is fixed. This keeps future s
 
 ### Tools quick reference
 
+The installed `memorix serve --mode lite` surface exposes every tool below
+except the three marked `MEMORIX_MODE=full` (set that env var or `--mode full`
+to enable them).
+
 | Tool | Use when |
 |---|---|
 | `memorix_search` | Find relevant past context |
@@ -73,9 +77,9 @@ Use `memorix_resolve` when a task is done or a bug is fixed. This keeps future s
 | `memorix_session_start` | Load session context (handoff, team coordination) |
 | `memorix_timeline` | See chronological context around a memory |
 | `memorix_retention` | Check memory health and archive expired items |
-| `memorix_promote` | Turn repeated patterns into permanent skills |
-| `memorix_rules_sync` | Inspect or sync rules across agents |
-| `memorix_workspace_sync` | Inspect or migrate workspace integrations |
+| `memorix_promote` (MEMORIX_MODE=full) | Turn repeated patterns into permanent skills |
+| `memorix_rules_sync` (MEMORIX_MODE=full) | Inspect or sync rules across agents |
+| `memorix_workspace_sync` (MEMORIX_MODE=full) | Inspect or migrate workspace integrations |
 
 ## Active Work
 

package/README.md

@@ -435,7 +435,7 @@ This opens memcode, a terminal coding agent that uses the same Memorix project m
 
 Search is project-scoped by default. `scope="global"` searches across projects. The search boosts Git Memory for "what changed" questions and reasoning records for "why" questions.
 
-Long-term memory is deliberately not an automatic dump of every note. A source observation, Claim, workflow, session, and code snapshot keep their existing roles. An agent may ask `memorix_store` to create an additional long-term **candidate**, but candidates never enter task context. Use `memorix memory long-term qualify|approve|archive|supersede` to record the evidence-backed lifecycle. Only a manually created or user-confirmed `user + portable` item may be considered in another local project; project code, Git facts, tests, workflows, sessions, and observations cannot be promoted into portable user memory.
+Long-term memory is deliberately not an automatic dump of every note. A source observation, Claim, workflow, session, and code snapshot keep their existing roles. An agent may ask `memorix_store` to create an additional long-term record; an explicit request auto-qualifies and enters task briefs as a durable anchor, while hook-captured or Git-derived candidates stay pending until qualified. Approval stays an explicit operator review. Use `memorix memory long-term approve|archive|supersede` to record the evidence-backed lifecycle. Only a manually created or user-confirmed `user + portable` item may be considered in another local project; project code, Git facts, tests, workflows, sessions, and observations cannot be promoted into portable user memory.
 
 `memorix context "..."` is the default Memory Autopilot entry. It builds a compact task-lensed brief for agents: bugfix tasks lean toward tests and repros, release tasks lean toward package/changelog/build checks, onboarding tasks lean toward docs and entry points, and stale or unrelated memories stay in warning lanes instead of flooding the prompt. A normal new task does not receive an old-session dump. For an explicit continuation, `memorix resume "..."` adds only the latest useful session summary, up to three readable durable anchors, and at most one recent source-labelled host compact checkpoint. A durable anchor carries a `durable:<id>` reference, so an agent can expand the full reviewed record through `memorix_detail` only when needed. Keyword matches stay primary; when no reviewed durable item matches and an embedding provider is configured, Memorix makes one 1.8-second, no-retry semantic fallback for paraphrases or cross-language tasks. A slow or unavailable provider simply leaves the normal keyword-only Workset intact. A checkpoint is lifecycle evidence, not durable memory or a transcript backup. Agents should read the suggested files before trusting stored memory.
 
@@ -450,12 +450,15 @@ Long-term memory is deliberately not an automatic dump of every note. A source o
 | Debug HTTP MCP in the foreground | `memorix serve-http --port 3211` |
 | Inspect or manage memory directly | `memorix memory`, `memorix reasoning`, `memorix session`, `memorix ingest`, `memorix media` |
 | Manage reviewed long-term memory | `memorix memory long-term list|show|add|promote|qualify|approve|archive|supersede` |
+| Retire memories from retrieval | `memorix purge` (current project), `memorix purge --all --yes` (everything) |
 | Inspect native compaction continuity | `memorix checkpoint list|show|context|archive` |
 | Use the interactive terminal memory control plane | `memorix workbench` |
 | Use the bundled terminal agent | `memorix` or `memcode` |
 | Run orchestrated subagent work | `memorix orchestrate --goal "..."` |
 
-`memorix serve` defaults to `--mode micro` (9 tools) to keep MCP tool schemas small for agents. Use `--mode lite` for the extended solo memory surface, `--mode team` for coordination tools, or `--mode full` for advanced and compatibility tools such as checkpoint inspection.
+`memorix serve` defaults to `--mode micro` (9 tools) to keep MCP tool schemas small for agents. `memorix setup` writes `--mode lite` into the installed MCP config so every tool taught in the generated guidance is reachable. Use `--mode team` for coordination tools, or `--mode full` for advanced and compatibility tools such as checkpoint inspection.
+
+Shortcuts and aliases: `memorix search` / `remember` / `recent` are direct query shortcuts, `bg`/`bs` start the background service, `hook` runs the hook launcher, and `memorix config get|path|migrate` inspects or migrates config files.
 
 `memorix orchestrate` uses the current checkout for single-worker runs. When running multiple workers, it creates task worktrees under `.worktrees/` and merges successful task branches back. Use `--isolated` to force worktree isolation for one worker, `--no-worktree` to disable it, `--allow-dirty` to run with uncommitted changes, and `--no-auto-merge` to preserve task worktrees for manual review.