v0.9.0 — the trust release

v0.9.0 — 2026-07-06

• Completed the Memory Operations Protocol — all ten verbs are real: merge via consolidation-candidate acceptance that executes member supersessions in one audited action; expire/unexpire riding the read-path validity exclusion; and true redact — content expunged from memories, revisions, and event payloads through a narrowly trigger-guarded redaction mode (append-only stays the default posture; the audit skeleton and a redaction proof-trail survive; migration 20260706_0079). All wired across MCP (alice_memory_manage), HTTP, and CLI with policy vocabulary (redact and consolidation-acceptance require human or admin).

• Context API v2: per-section token allocation in the budget report, five packing strategies (balanced/facts_first/recent_first/contradictions_first/sources_first), and deterministic depth tiers (minimal/low/medium/high — no tier performs model synthesis); tri-state include flags let tier defaults breathe; the default agent loop docs now center one context call.

• Complete export/import round-trip: export now covers all nine record types (entities, edges, revisions, provenance, and chunks were previously dropped); alice-memory import preserves ids and timestamps exactly, never overwrites, and is all-or-nothing.

• Published the scale envelope (docs/benchmarks/scale/): SQLite commits flat at 2.3ms through 100k memories after this benchmark caught and fixed a 300× O(N) idempotency scan; Postgres ~20ms commits and ~400ms recall at 100k; honest SQLite-with-embeddings boundary documented.

• Entity-extraction hygiene after a LongMemEval diagnostic: bare capitalized spans no longer default to person (positive evidence required), long-text repeat thresholds and confidence-ranked caps stop conversational noise flooding; extraction rule + confidence recorded per entity for future re-typing.

• LongMemEval documentation: three-run variance disclosure (≈64%, band 63.0–64.6), the disclosed negative result on entity-graph retrieval for multi-session, and a breadth ablation (49.2% multi-session at 2× context) motivating the planned aggregation mode.

• Temporal graph memory + entity resolution (Sprint D): a generic vnext_entities substrate with canonicalization, aliases, mention windows, and append-only relationship history (migration 20260705_0078); deterministic entity extraction (capitalized spans, acronyms, handles, domains, repeat-thresholds, blocklist — no LLM) linking sources at capture and memories at acceptance on every acceptance path; entity-hop graph retrieval fused into RRF as a third stage with full trace honesty; a belief-evolution timeline in alice_explain; and two new eval suites — entity_resolution and graph_hop_retrieval — where the graph mechanism proves recall 1.0 on entity-only queries that lexical search scores 0.0 on.

Pre-launch fixes

• Human-direct memory commits no longer require an agent identity via MCP.
• SQLite MCP server bootstraps the user row automatically (python -m path) with clearer integrity-error messages.
• Full-text recall falls back to OR-matching when strict AND finds nothing (the trace shows the fallback).
• CLI gains --version, friendly errors for sqlite URLs and bad UUIDs, and lists all six eval suites.
• Docs overhaul: pip/uvx install is the primary quickstart, the eleven-core-tool count is corrected everywhere, self-host role bootstrap SQL is documented, and PyPI metadata is completed.

Alice v0.8.0 — the memory-frontier release

Alice v0.8.0 Release Notes

v0.8.0 is the memory-frontier release: the release where Alice's vocabulary
started doing real work, and the release that ships Alice's first published
benchmark result.

Benchmark

64.6% on LongMemEval_s (323/500, official judge protocol, GPT-4o
generation, hybrid retrieval with embeddings) — in the same range as the best
published results in the category. Full methodology, per-question evidence,
and a reproduction script live in
docs/benchmarks/longmemeval/.
Knowledge-update questions score 74.4% — the correction/supersession
machinery measurably working. The weakest category, multi-session synthesis
(45.1%), is the top roadmap item.

Highlights

Context packs that respect budgets

• max_tokens is enforced with greedy budget packing; packs report
  {token_budget, token_estimate, truncated, dropped_item_count}.
• Contradicting evidence and recent changes are real sections; stale
  sections removed; rows no longer duplicated.

Typed, staleness-aware retrieval

• memory_types and projects filters end-to-end (stores → compiler →
  alice_recall/alice_context_pack).
• Procedure:/Playbook:/How to and Happened:/Log: capture rules.
• Expired facts (valid_to < now) excluded from search by default; stale
  status; daily staleness_sweep workflow marks — never deletes — expired
  and long-unconfirmed volatile memories for review.

The agentic write protocol, on the front door

• alice_memory_commit and alice_memory_manage (confirm/undo/forget) join
  the core MCP surface — 11 tools, every parameter described.
• The Memory Operations Protocol
  documents every verb with its audit guarantees and honest boundaries.

Real scopes

• project_id, created_by_agent_id, run_id columns with metadata
  backfill; filters everywhere; agent API keys can bind a project scope —
  bound identities may narrow but never widen it.

Consolidation that consolidates

• Embedding-based near-duplicate clustering produces merge/dedup candidates
  through the review gate. Model-backed merges are grounding-gated with
  structured refusals; the deterministic path never fabricates text;
  supersession is never automatic.

Temporal slice

• Graph edges carry event time; supersession pointers are first-class
  columns; as-of edge queries on both backends; alice_explain walks the
  full supersession chain; the SQLite on-ramp gains the graph substrate.

Evals

• Four live suites (retrieval_quality, correction_suppression,
  decision_recovery, provenance_explanation) run the production pipeline
  on both backends and can genuinely fail.
• The LongMemEval harness ships in-repo (eval/longmemeval/).

Upgrading from v0.7.0

1. alembic -c apps/api/alembic.ini upgrade head (migrations 0075 → 0077:
   staleness indexes, scope columns with backfill, temporal columns with
   backfill — all additive and reversible).
2. SQLite on-ramp files upgrade in place automatically at next start.
3. MCP clients see two new core tools and new optional arguments; the nine
   existing tools keep their names and required arguments.
4. Optional: enable the staleness_sweep scheduler workflow; create agent
   keys with --project-scope to bind agents to projects.

Verification

1,460 unit and 377/377 integration tests green; live eval battery 4/4;
migration chain verified; the benchmark run itself exercised the full
capture → retrieval → generation pipeline 500 times against this code.

Alice v0.7.0 — zero-infrastructure on-ramp

The first real alice-memory release on PyPI. Try Alice with nothing but Python:

uvx alice-memory mcp --data-dir ~/.alice

One command, no Docker, no Postgres: a working MCP memory server over a local SQLite file, exposing the nine core tools — capture with review-governed approval, hybrid recall (FTS5 full-text now, embedding-based vector search the moment you configure an OpenAI-compatible endpoint), resumption briefs, recent decisions, open loops, corrections with audit trails, and provenance explanations.

Highlights

• SQLite backend mirroring the vNext store surface: FTS5 with porter stemming, numpy cosine vector search, explicit per-user scoping, append-only audit trails.
• alice-memory CLI: mcp (serve stdio MCP) and export (JSONL dump of memories, sources, open loops, events).
• Same tools, same semantics as Postgres — swap the database URL when you outgrow the trial; Postgres remains the recommended path for full workflows (web review console, scheduler, legacy continuity surfaces).
• CI-benchmarked retrieval: the live retrieval-quality suite now runs against in-memory SQLite (lexical recall@1 = 1.0 through the production pipeline, ~0.7 ms median per query).
• Published via PyPI Trusted Publishing — no API tokens in the release path.

Full details in CHANGELOG.md and docs/release/v0.6.0-release-notes.md for the v0.6.0 overhaul this builds on.

Alice v0.6.0 — the continuity layer for AI agents

Alice v0.6.0 Release Notes

v0.6.0 is the product-viability overhaul release. It repositions Alice as
the continuity layer for AI agents and replaces the weakest parts of the
previous baseline with production-grade implementations.

Highlights

Hybrid retrieval (new)

• Memory search runs Postgres full-text and pgvector (HNSW) similarity, fused
  with reciprocal-rank fusion.
• Embeddings come from any OpenAI-compatible endpoint (Ollama, LM Studio,
  OpenAI) via ALICE_EMBEDDINGS_BASE_URL / ALICE_EMBEDDINGS_MODEL /
  ALICE_EMBEDDINGS_API_KEY. Memories embed at write time.
• Without an embedding endpoint, search degrades to full-text only and states
  that explicitly in the retrieval trace.
• Contradiction sync no longer runs inside the read path.

Nine-tool MCP surface (changed)

• The default tools/list exposes nine core tools: alice_capture,
  alice_recall, alice_resume, alice_context_pack, alice_open_loops,
  alice_recent_decisions, alice_memory_review, alice_memory_correct,
  alice_explain — every parameter documented, outputs compact.
• The legacy long-tail surface (65 tools, including all alice_vnext_* names)
  remains available with ALICE_MCP_LEGACY_TOOLS=1. The legacy continuity
  recall view is alice_recall_debug.

Per-agent API keys (new)

• Create keys with alicebot agent keys create --agent-id <id> --profile <profile>;
  only the sha256 hash is stored, scoped by row-level security.
• All vNext agent HTTP endpoints accept Authorization: Bearer alice_sk_...;
  identity and permission profile come from the key record, and privilege
  escalation attempts are rejected and audited.
• Keyless local setups keep working until the first key is created.
• The MCP server binds a key via the ALICE_AGENT_API_KEY environment variable.

Honest evals (changed)

• The previous eval suites, which verified their own fixtures, are removed.
• The retrieval_quality benchmark seeds memories through the real store and
  measures recall@k, MRR, and latency through the actual retrieval pipeline.
  Live runs require ALICEBOT_EVAL_DATABASE_URL; without it, suites report
  skipped, never a fabricated pass. See eval/README.md.

Fixes

• Scheduler bootstrap no longer crashes on live Postgres stores: the
  memory_consolidation workflow type is now allowed by the scheduler check
  constraints (migration 20260704_0074).
• Alembic prefers an explicitly passed database URL over the
  DATABASE_ADMIN_URL / DATABASE_URL environment variables, so per-test
  databases migrate correctly.

Upgrading from v0.5.1

1. Pull v0.6.0 and run alembic -c apps/api/alembic.ini upgrade head
   (adds retrieval columns/indexes, agent_api_keys, and the scheduler
   constraint fix).
2. Optionally configure an embedding endpoint for semantic search.
3. MCP clients that used legacy tools set ALICE_MCP_LEGACY_TOOLS=1, or
   migrate to the nine core tools.
4. To authenticate agents, create keys and pass them as Bearer tokens; nothing
   changes until the first key exists.

Verification

• 1,249 unit tests and 377/377 integration tests green (integration verified
  with proper role separation: alicebot_app for app connections, admin for
  migrations).
• Full migration chain verified up → down → up on a fresh pgvector database.
• Live retrieval_quality run: lexical recall@1 = 1.0 at ~1.5 ms median
  per query (FTS-only mode); the suite demonstrably fails under corrupted
  vectors.

Packaging

• The Python package name on PyPI is
  alice-memory (0.0.1 placeholder;
  the packaged runtime ships under this name). alice-core on PyPI is an
  unrelated project.
• This release still runs from a repo checkout:
  make setup && make migrate && make doctor && make dev.

Known boundaries

Local-first and single-user; no hosted service, no OAuth connectors, no
automatic capture from arbitrary conversation. See
docs/alpha/known-limitations.md.

Alice v0.6.0-alpha-rc.3

v0.6.0-alpha-rc.3 Release Notes

Status: internal pre-release candidate for local alpha and Hermes dogfood testing.

Latest stable remains: v0.5.1.

Do not mark this release as latest. Publish it as a GitHub pre-release and not latest with --latest=false.

Purpose

v0.6.0-alpha-rc.3 supersedes v0.6.0-alpha-rc.2. rc.2 remains immutable; rc.3 carries the latest current-main local alpha fixes, agentic memory commit workflow, retrieval verification, MCP error hardening, and installer/check updates.

The target dogfood setup remains Alice and Hermes running locally or on the same Ubuntu server/VPS, with Alice accessed through localhost or SSH tunneling and Hermes connected through MCP/API.

Changes Since rc.2

• Added governed agentic memory commit support across API, CLI, MCP tools, Hermes/OpenClaw skills, audit/revision trails, correction, undo, forget, recent commits, and policy enforcement.
• Hardened MCP unexpected-error handling so failed memory operations return structured tool errors without closing the MCP connection.
• Fixed Postgres-backed vNext retrieval so keyword-style context-pack queries can select matching committed memories and sources even when the full query is not one contiguous phrase.
• Verified retrieval for domain=unknown committed memories when agents request scoped packs such as professional, project, and personal.
• Fixed local /vnext?mode=live CORS configuration for alpha/dev by requiring explicit localhost origins instead of weakening production defaults.
• Updated local env templates, Ubuntu installer config generation, doctor checks, troubleshooting docs, and local-cors smoke coverage.
• Documented the clean web dev cache reset path for stale .next chunk errors.

Included

• Headless Ubuntu installer: scripts/install-ubuntu.sh
• Safe uninstall/reset helper: scripts/uninstall-ubuntu.sh
• Systemd templates for alice-api, alice-web, and alice-scheduler
• Local config template at packaging/ubuntu/alicebot.env.example
• Public alpha docs and agent integration docs under docs/alpha/
• Agent skill docs under agent-skills/
• Agentic memory commit smoke: alicebot vnext smoke agentic-memory-commit
• Local CORS readiness and smoke checks
• Existing connector, operator-console, capture-to-brief, scheduler, and agent integration smokes

Security Defaults

• API and web bind to localhost by default for local alpha.
• /vnext local live mode requires explicit CORS origins; wildcard CORS is not introduced.
• Secrets are not printed by the installer or readiness checks.
• Agent memory writes are governed by permission profile, domain, sensitivity, confirmation, and review policy.
• Read-only agents cannot write memory.
• Project-scoped agents remain constrained to their allowed project/domain context.
• Sensitive memory requires confirmation or review before becoming durable context.

Install Command

Inspect before running:

curl -fsSL https://raw.githubusercontent.com/samrusani/AliceBot/v0.6.0-alpha-rc.3/scripts/install-ubuntu.sh -o install-alice.sh
less install-alice.sh
bash install-alice.sh --tag v0.6.0-alpha-rc.3

For headless access through SSH:

ssh -L 3000:127.0.0.1:3000 -L 8000:127.0.0.1:8000 user@ubuntu-box

Validation

• ./.venv/bin/pytest tests/unit -q -> 1142 passed
• ./.venv/bin/pytest tests/integration -q -> 377 passed
• pnpm --dir apps/web test -> 209 passed
• pnpm --dir apps/web lint -> passed
• pnpm --dir apps/web build -> passed
• ./.venv/bin/alicebot eval run --suite all -> pass, 170/170 cases, 0 critical privacy leaks, 0 prompt-injection tool writes
• ./.venv/bin/alicebot vnext alpha check -> passed
• ./.venv/bin/alicebot vnext smoke agentic-memory-commit -> passed
• ./.venv/bin/alicebot vnext smoke agent-integration-pack -> passed
• ./.venv/bin/alicebot vnext smoke operator-console -> passed
• ./.venv/bin/alicebot vnext smoke capture-to-brief -> passed
• ./.venv/bin/alicebot vnext smoke agentic-scheduler -> passed
• git diff --check -> passed
• GitHub Security Scans on current main c2fa8e0e7fe9a6ae106fa4c1b56163d41078143e -> Gitleaks, CodeQL Python, and CodeQL JavaScript passed
• Protected Path Guardrails for PR #219, which produced current main, passed

Release Assets

• install-ubuntu.sh
• SHA256SUMS

GitHub source archives are generated automatically for the tag.

Alice v0.6.0-alpha-rc.2

v0.6.0-alpha-rc.2 Release Notes

Status: internal pre-release candidate for Hermes dogfood testing.

Latest stable remains: v0.5.1.

Do not mark this release as latest. Publish it as a GitHub pre-release and not latest with --latest=false.

Purpose

v0.6.0-alpha-rc.2 supersedes v0.6.0-alpha-rc.1 for headless Ubuntu installs. The original RC remains immutable, and this RC carries the same headless Ubuntu/Hermes dogfood scope with additional installer hardening from the first Ubuntu install attempt.

The target dogfood setup is Alice and Hermes running on the same Ubuntu server or VPS, with Alice accessed through SSH tunneling and Hermes connected through MCP/API.

Included

• headless Ubuntu installer: scripts/install-ubuntu.sh
• safe uninstall/reset helper: scripts/uninstall-ubuntu.sh
• systemd templates for alice-api, alice-web, and alice-scheduler
• local config template at packaging/ubuntu/alicebot.env.example
• headless install guide at docs/alpha/headless-ubuntu-install.md
• Hermes dogfood guide at docs/alpha/hermes-dogfood-ubuntu.md
• headless package smoke: alicebot vnext smoke headless-ubuntu
• headless alpha readiness mode: alicebot vnext alpha check --headless
• public alpha agent integration pack from the prior packaging phase

Installer Hardening Since rc.1

• make setup now creates .env and .env.lite when missing.
• Environment files are validated before startup scripts source them.
• Values with spaces, including ALICE_MCP_COMMAND, are quoted in the Ubuntu template.
• Local Ubuntu installs default to APP_ENV=development instead of production.
• Startup scripts require the project venv and fail with a setup hint instead of silently falling back to system Python.
• pnpm is pinned to 10.23.0, approved native build dependencies are declared, and the web install uses the repo wrapper.
• Docker Compose Postgres credentials can be driven from the active Alice env.
• The Postgres init role script creates or updates the app role with the configured password.

Security Defaults

• API and web bind to 127.0.0.1 by default.
• /vnext is intended to be accessed through SSH tunnel first.
• The installer preserves existing config and does not overwrite secrets.
• Generated local database passwords are not printed.
• Agent outputs and memory proposals remain review-only.
• Trusted memory is not auto-promoted.

Known Limitations

• This is not hosted cloud deployment.
• This is not public SaaS auth.
• This is not a public beta.
• Gmail OAuth, Calendar OAuth, OCR execution, voice transcription, mobile app, billing, and team accounts remain out of scope.
• The installer is documented for Ubuntu 22.04 and 24.04.
• Systemd templates are local-alpha templates and should be reviewed before production hardening.

Recommended GitHub Release Command

git tag -a v0.6.0-alpha-rc.2 -m "v0.6.0-alpha-rc.2"
git push origin v0.6.0-alpha-rc.2
gh release create v0.6.0-alpha-rc.2 \
  --title "Alice v0.6.0-alpha-rc.2" \
  --notes-file docs/release/v0.6.0-alpha-rc.2-release-notes.md \
  --prerelease \
  --latest=false

Optional assets:

git archive --format=tar.gz --prefix=AliceBot-v0.6.0-alpha-rc.2/ \
  -o alicebot-v0.6.0-alpha-rc.2-source.tar.gz v0.6.0-alpha-rc.2
sha256sum alicebot-v0.6.0-alpha-rc.2-source.tar.gz scripts/install-ubuntu.sh > SHA256SUMS
gh release upload v0.6.0-alpha-rc.2 alicebot-v0.6.0-alpha-rc.2-source.tar.gz SHA256SUMS scripts/install-ubuntu.sh

Install Command

Inspect before running:

curl -fsSL https://raw.githubusercontent.com/samrusani/AliceBot/v0.6.0-alpha-rc.2/scripts/install-ubuntu.sh -o install-alice.sh
less install-alice.sh
bash install-alice.sh --tag v0.6.0-alpha-rc.2

Then access /vnext through:

ssh -L 3000:127.0.0.1:3000 -L 8000:127.0.0.1:8000 user@ubuntu-box

Validation

• bash -n for installer, runtime startup scripts, env validator, pnpm install wrapper, and Postgres init script
• ./scripts/validate_env.sh .env .env.lite.example packaging/ubuntu/alicebot.env.example
• docker compose config
• make setup
• make migrate
• make doctor
• ./.venv/bin/pytest tests/unit -q
• pnpm --dir apps/web test
• pnpm --dir apps/web lint
• pnpm --dir apps/web build
• ./.venv/bin/alicebot vnext smoke headless-ubuntu
• ./.venv/bin/alicebot vnext alpha check --headless --skip-smokes
• Ubuntu 24.04 installer dry-run through Docker
• git diff --check

v0.6.0-alpha-rc.1

v0.6.0-alpha-rc.1 Release Notes

Status: internal pre-release candidate for Hermes dogfood testing.

Latest stable remains: v0.5.1.

Do not mark this release as latest. If published on GitHub, publish it as a pre-release and not latest.

Purpose

v0.6.0-alpha-rc.1 packages Alice vNext for a headless Ubuntu install path from GitHub. The target dogfood setup is Alice and Hermes running on the same Ubuntu server or VPS, with Alice accessed through SSH tunneling and Hermes connected through MCP/API.

This headless Ubuntu RC is intended for operator-controlled dogfood only.

Included

• headless Ubuntu installer: scripts/install-ubuntu.sh
• safe uninstall/reset helper: scripts/uninstall-ubuntu.sh
• systemd templates for alice-api, alice-web, and alice-scheduler
• local config template at packaging/ubuntu/alicebot.env.example
• headless install guide at docs/alpha/headless-ubuntu-install.md
• Hermes dogfood guide at docs/alpha/hermes-dogfood-ubuntu.md
• headless package smoke: alicebot vnext smoke headless-ubuntu
• headless alpha readiness mode: alicebot vnext alpha check --headless
• public alpha agent integration pack from the prior packaging phase

Security Defaults

• API and web bind to 127.0.0.1 by default.
• /vnext is intended to be accessed through SSH tunnel first.
• The installer preserves existing config and does not overwrite secrets.
• Generated local database passwords are not printed.
• Agent outputs and memory proposals remain review-only.
• Trusted memory is not auto-promoted.

Known Limitations

• This is not hosted cloud deployment.
• This is not public SaaS auth.
• This is not a public beta.
• Gmail OAuth, Calendar OAuth, OCR execution, voice transcription, mobile app, billing, and team accounts remain out of scope.
• The installer is documented for Ubuntu 22.04 and 24.04.
• Systemd templates are local-alpha templates and should be reviewed before production hardening.

Recommended GitHub Release Command

git tag -a v0.6.0-alpha-rc.1 -m "v0.6.0-alpha-rc.1"
git push origin v0.6.0-alpha-rc.1
gh release create v0.6.0-alpha-rc.1 \
  --title "Alice v0.6.0-alpha-rc.1" \
  --notes-file docs/release/v0.6.0-alpha-rc.1-release-notes.md \
  --prerelease \
  --latest=false

Optional assets:

git archive --format=tar.gz --prefix=AliceBot-v0.6.0-alpha-rc.1/ \
  -o alicebot-v0.6.0-alpha-rc.1-source.tar.gz v0.6.0-alpha-rc.1
sha256sum alicebot-v0.6.0-alpha-rc.1-source.tar.gz scripts/install-ubuntu.sh > SHA256SUMS
gh release upload v0.6.0-alpha-rc.1 alicebot-v0.6.0-alpha-rc.1-source.tar.gz SHA256SUMS scripts/install-ubuntu.sh

Install Command

Inspect before running:

curl -fsSL https://raw.githubusercontent.com/samrusani/AliceBot/v0.6.0-alpha-rc.1/scripts/install-ubuntu.sh -o install-alice.sh
less install-alice.sh
bash install-alice.sh --tag v0.6.0-alpha-rc.1

Then access /vnext through:

ssh -L 3000:127.0.0.1:3000 -L 8000:127.0.0.1:8000 user@ubuntu-box

Alice v0.5.1 vNext Preview

v0.5.1 vNext Preview Release Notes

Release date: 2026-05-11
Preview tag: v0.5.1-vnext-preview
Stable public release: v0.5.1
Release type: public pre-release preview

Summary

Alice vNext is now ready as a public preview of the true second-brain direction on top of the stable v0.5.1 baseline. The preview is local-first and deterministic: it proves the vNext memory kernel, reviewable generated artifacts, connector-backed evidence capture, context packs, CLI/API/MCP access, eval harness, and fixture-backed /vnext workspace without claiming hosted SLA or live connector automation.

Included Preview Surface

• vNext memory-kernel schema, shared JSON contracts, Postgres store facade, event log, and Brain Charter model.
• Source capture for manual text, local text/Markdown files, Markdown folders, and ChatGPT exports.
• Context packs with domain/sensitivity filters, provenance, retrieval trace metadata, CLI/API/MCP access, and JSON-safe Postgres row handling.
• Queue/artifact workflows with reviewable generated artifacts, export actions, and explicit review state.
• Daily brief, weekly synthesis, connection report, contradiction/belief, project update, and open-loop workflows.
• Deterministic connector payload ingestion for Telegram, browser clipper, PDF, DOCX, CSV, screenshot OCR, and voice transcript payloads.
• Fixture-backed /vnext workspace covering Home, Inbox, Ask Alice, briefs, queue, generated artifacts, memory review, projects, people, beliefs, open loops, timeline, graph, and settings/privacy.
• Synthetic eval corpus and baseline suites for recall, temporal reasoning, contradictions, provenance, privacy, open loops, and prompt-injection write safety.
• Public-preview docs: quickstart, architecture, security/privacy, contributor guide, synthetic demo dataset, demo video script, release checklist, and tag plan.

Verification

Release-readiness evidence recorded on 2026-05-11:

• Real Postgres vNext smoke passed for source capture, connector ingest, scoped and unscoped context packs, daily brief generation, project update review, open-loop close, API context packs, API project dashboard, MCP context pack, and MCP project dashboard.
• ./.venv/bin/python -m pytest tests/unit -q: 1057 passed.
• pnpm --dir apps/web test: 205 passed.
• pnpm --dir apps/web lint: passed.
• pnpm --dir apps/web build: passed and built /vnext.
• python3 scripts/check_control_doc_truth.py: passed.
• ./.venv/bin/python -c 'from alicebot_api.cli import main; raise SystemExit(main(["eval", "run", "--suite", "all"]))': 170/170 cases, zero critical privacy leaks, zero prompt-injection tool writes.
• git diff --check: clean.
• GitHub Security Scans on merged main passed for CodeQL JavaScript, CodeQL Python, and Gitleaks.

Known Limitations

• vNext is a local-first preview, not a hosted launch.
• Connector support is deterministic payload ingestion only; live OAuth, external polling, browser extension actions, OCR model execution, and transcription model execution are deferred.
• Generated artifacts are reviewable and are not auto-promoted to trusted memory.
• Production daily/weekly scheduling is deferred.
• The /vnext workspace is fixture-backed; live-backed UI expansion remains future work.
• Model-backed or human-rated eval scoring is deferred beyond this deterministic synthetic preview.

Upgrade And Rollback

Use v0.5.1 for the stable pre-1.0 public release line. Use v0.5.1-vnext-preview only when testing the preview surface. Roll back by checking out v0.5.1 or by disabling vNext routes and avoiding the 20260510_0067_vnext_memory_kernel_schema.py migration in a disposable preview database.

Alice v0.5.1: Provider Adapters + Design Partner Launch

Alice v0.5.1 is the Phase 14 platform release for Alice.

This release makes Alice easier to plug into local, self-hosted, enterprise-curious, and external-agent workflows without changing its core continuity semantics.

Highlights

• Provider/runtime portability across OpenAI-compatible endpoints, Ollama, llama.cpp, vLLM, and Azure-backed paths.
• First-party model packs for llama, qwen, gemma, and gpt-oss with provider-aware bindings and pack-aware defaults.
• Reference integration paths for Hermes, OpenClaw, and generic Python and TypeScript agents.
• Design-partner onboarding, workspace linkage, usage summaries, feedback intake, and launch evidence.
• Logging-safety hardening from HF-001: stdout-by-default local/Lite logging, disabled local/Lite access logs by default, and bounded opt-in file logging.

What shipped in this boundary

• P14-S1: Provider abstraction cleanup + OpenAI-compatible adapter
• P14-S2: Ollama + llama.cpp + vLLM adapters
• P14-S3: Model packs
• P14-S4: Reference integrations
• P14-S5: Design partner launch
• HF-001: Logging safety and disk guardrails

Release posture

• Pre-1.0 public release
• Version boundary: v0.5.1
• Main focus: platform compatibility, adoption, and operational safety

For release verification and operational details, see the repo changelog and release docs in docs/release/.

v0.4.0

Alice v0.4.0 makes the system easier to adopt without changing its core continuity model.

This release adds a single continuity entrypoint for external agents, a lighter local startup path, and clearer visibility into memory quality and conversation health.

What is included

• One-call continuity across API, CLI, and MCP
• Alice Lite for a faster local first run
• Memory hygiene visibility for duplicates, stale facts, contradictions, weak trust, and review pressure
• Conversation health visibility for recent, stale, and risky threads

Why this matters

• External agents can integrate through one main continuity call instead of stitching together multiple surfaces
• Solo users and builders can get Alice running locally with less setup friction
• Operators can see quality and risk posture more clearly without weakening Alice's continuity semantics

Baseline included in this release

• Phase 12 retrieval, mutation, trust calibration, eval, and briefing stack
• Hermes bridge support
• Provider runtime and model-pack support

Verification

• 1279 Python tests passed
• 200 web tests passed
• Alice Lite smoke passed
• Hermes provider, MCP, and bridge demo smoke passed
• Public eval harness passed 12/12 cases

Release boundary
This is the public pre-1.0 release for the completed Phase 13 surface.