Persistent Memory 1.2.0 — Obsidian and Porcelain themes

GitHub release v1.2.0

Service	Version	Change
dashboard / dashboard layer	1.2.0	Neural Obsidian design, Obsidian and Porcelain themes, persistent appearance controls and accessible overlays.
onboarding	1.2.0	Matching wizard design, responsive layout, keyboard theme controls and dashboard appearance handoff.
docs	1.2.0	Refreshed interface screenshots and appearance guidance.
api	1.0.2	Existing personal memory and provider health APIs retained.
update-runner / update-flow layer / update-coordinator	1.0.1	Existing published-release discovery, pinned upgrades and recovery retained.
shared / schema	1.1.0	Existing embedding catalog and schema retained.
worker / database / mcp / mcp-runtime / core-tools	1.0.0	Existing processing, storage and memory tools retained.
graph / graphiti service / memory-vector / evidence-files / security-dlp / DLP service	1.0.0	Existing graph, retrieval and sensitive-data checks retained.
dashboard-gateway / docker-control	1.0.0	Existing local gateway and service controls retained.

• Refresh the dashboard and installation wizard with the Neural Obsidian design:
  cyan-to-violet accents, clearer surface hierarchy and locally bundled fonts.
• Choose the default dark Obsidian or light Porcelain theme. The dashboard also
  follows the operating system through Match system, including after appearance
  controls close. Theme changes synchronize across open dashboard tabs.
• Carry the wizard's selected appearance into the dashboard, even when browser
  storage is unavailable. Theme controls support arrow keys, Home and End.
• Fix profile and navigation dialogs appearing behind page content. Improve
  light-theme notices, status badges, image-preview controls, graph retry controls
  and small-screen wizard layout. Keep graph legend colors aligned with nodes.
• Avoid memory-list hydration errors when Docker and the browser use different
  time zones. Redact passwords embedded in installer-preview URLs and Neo4j
  credentials without changing the saved environment.
• Fix the documentation landing redirect so its styles, images and navigation
  remain inside the dashboard's documentation route.
• Refresh public screenshots using generic demonstration data and document
  appearance controls. The personal installation wizard has eleven steps;
  optional connections remain a later dashboard action.

Direct upgrades from public 1.0.0 through 1.1.4 preserve memories, data volumes,
credentials and embedding configuration. This release adds no database migration
or embedding-model change. The graph canvas retains its dark backdrop in both themes.

To update an existing installation, run from its repository directory:

npm run update-persistent-memory -- --release 1.2.0

Windows PowerShell: npm.cmd run update-persistent-memory -- --release 1.2.0.
For a new installation, follow the
v1.2.0 installation instructions.

Validation covers theme persistence, OS and browser-tab synchronization, keyboard
controls, restricted browser storage, appearance handoff, dashboard interactions
and the personal wizard flow. Physical macOS/Safari installation was not rerun
on this Windows development machine.

Persistent Memory 1.1.4 — API-only installation and cleanup fixes

GitHub release v1.1.4

Service	Version	Change
dashboard / dashboard layer	1.1.4	Current release label and release notes.
onboarding	1.1.3	Required-runtime startup grace and visible verification diagnostics.
docs	1.1.4	Optional Ollama monitoring, verification troubleshooting and scoped dependency-image cleanup.
api	1.0.2	Monitor only required host Ollama, retain configured model names before first use and count unhealthy local models correctly.
update-runner / update-flow layer / update-coordinator	1.0.1	Existing published-release discovery, pinned upgrades and recovery retained.
shared / schema	1.1.0	Existing embedding catalog and schema retained.
worker / database / mcp / mcp-runtime / core-tools	1.0.0	Existing processing, storage and memory tools retained.
graph / graphiti service / memory-vector / evidence-files / security-dlp / DLP service	1.0.0	Existing graph, retrieval and sensitive-data checks retained.
dashboard-gateway / docker-control	1.0.0	Existing local gateway and service controls retained.

• API-only installations no longer probe or display an unused host Ollama
  service, count it as a dashboard failure, or project a historical host error.
  Server-managed local embeddings still check the host and selected model.
  Client-managed embeddings retain their own client-scoped health.
• Show the configured Fact extraction and Embeddings model names before their
  first observed request or test. Their health remains unknown until observed.
  A reachable Ollama host with a missing required model counts as unhealthy.
• Allow up to 120 seconds for running required containers to finish their initial
  healthchecks during final verification. Image and runtime validation failures now include
  their diagnostic output instead of only a generic failure summary.
• The optional uninstall image-removal step includes exact configured downloaded
  dependencies, owned application images and explicitly listed Alpine utility
  images (alpine:3.20 and alpine:latest). Preserve images referenced
  by any running or stopped container, images carrying unrelated ownership or
  extra tags, and unattributed base/cache images. Report why candidate images
  are retained; never force-remove them or run a global Docker prune.

Direct upgrades from public 1.0.0 through 1.1.3 preserve memories, data volumes,
credentials and embedding configuration. This patch adds no database migration
or embedding-model change.

To update an existing installation, run from its repository directory:

npm run update-persistent-memory -- --release 1.1.4

Windows PowerShell: npm.cmd run update-persistent-memory -- --release 1.1.4.
For a new installation, follow the
v1.1.4 installation instructions.

Regression coverage exercises API-only and local-model monitoring, startup
readiness, diagnostic propagation, and image-cleanup ownership safeguards.
The original Mac verification failure cannot be attributed to Ollama from the
reported summary: its embedding check already passed, and the failing check
was required image identity/runtime validation. A physical Mac reinstall has
not been verified.

Persistent Memory 1.1.3 — Reliable installation and instruction recovery

GitHub release v1.1.3

Service	Version	Change
dashboard / dashboard layer	1.1.3	Current release label and release notes.
onboarding	1.1.2	Safe instruction recovery, personal-only public wizard, supported Node checks and clearer setup failures.
docs	1.1.3	Node LTS selection, setup troubleshooting, instruction backups and the personal installation flow.
api	1.0.1	Existing memory APIs and published-release metadata retained.
update-runner / update-flow layer / update-coordinator	1.0.1	Existing published-release discovery, pinned upgrades and recovery retained.
shared / schema	1.1.0	Existing embedding catalog and schema retained.
worker / database / mcp / mcp-runtime / core-tools	1.0.0	Existing processing, storage and memory tools retained.
graph / graphiti service / memory-vector / evidence-files / security-dlp / DLP service	1.0.0	Existing graph, retrieval and sensitive-data checks retained.
dashboard-gateway / docker-control	1.0.0	Existing local gateway and service controls retained.

• Recover incomplete or nested managed markers in existing Claude/Codex
  instruction files. Back up damaged originals byte-for-byte before writing,
  preserve their text outside the new managed block, and report each backup's
  location. Retries keep the recovered text and do not repeat the repair.
  Invalid custom blocks, unreadable encodings and failed backups stop safely.
• Remove Shared Memories from the public installation wizard. Review now leads
  directly to installation of the local Personal Memories stack, and hidden
  connection values cannot activate a shared connection. Existing operator
  connector flows and dashboard connections remain available separately.
• Require Node 24 LTS (24.x) or Node 22.12+ within 22.x for host installation.
  Check the running interpreter, keep macOS/Linux npm children on that same
  runtime, and explain when a successful Node installation requires restarting
  the wizard. Newer Current Node releases are not yet validated for this app.
• Name each dependency/setup substep before it runs and identify the failed
  substep without echoing private error details in the summary. This separates
  snapshot, dependency, Prisma, build and agent-registration failures.

Direct upgrades from public 1.0.0 through 1.1.2 preserve memories, data volumes,
credentials and embedding configuration. This patch adds no database migration
or embedding-model change. Use a supported Node LTS version before updating.

To update an existing installation, run from its repository directory:

npm run update-persistent-memory -- --release 1.1.3

Windows PowerShell: npm.cmd run update-persistent-memory -- --release 1.1.3.
For a new installation, follow the
v1.1.3 installation instructions.

Validation covers automated host, installer, instruction-recovery and release
checks, production installer builds and Chrome wizard checks. macOS runtime
selection is covered by controlled fixtures; a physical Mac installation has
not been verified. The reported installation failure was traced to incomplete
instruction markers. Node 26 was a separate runtime-policy gap, not evidence
of Prisma incompatibility.

Persistent Memory 1.1.2 — macOS setup and embedding diagnostics

GitHub release v1.1.2

Service	Version	Change
dashboard / dashboard layer	1.1.2	Current release label and release notes.
onboarding	1.1.1	macOS available-memory estimates and clearer, sanitized OpenAI embedding-test diagnostics.
docs	1.1.2	Explains macOS memory estimates, conservative fallbacks, and OpenAI request permissions.
api	1.0.1	Existing memory APIs and published-release metadata retained.
update-runner / update-flow layer / update-coordinator	1.0.1	Existing published-release discovery, pinned upgrades and recovery retained.
shared / schema	1.1.0	Existing embedding catalog and schema retained.
worker / database / mcp / mcp-runtime / core-tools	1.0.0	Existing processing, storage and memory tools retained.
graph / graphiti service / memory-vector / evidence-files / security-dlp / DLP service	1.0.0	Existing graph, retrieval and sensitive-data checks retained.
dashboard-gateway / docker-control	1.0.0	Existing local gateway and service controls retained.

• Fix false low-memory blocks on macOS by using an available-memory estimate
  from the free, speculative and inactive page counts reported by vm_stat.
  Honor its 4 KiB or 16 KiB page size and avoid adding overlapping purgeable
  pages. Preserve the raw free-memory measurement and retain all minimum
  requirements and local-model restrictions.
• Explain that inactive memory may require compression or disk writeback;
  the estimate is not a guarantee of immediately free RAM. Invalid counters,
  unavailable commands and timeouts trigger a visible, conservative fallback.
• Make OpenAI embedding-test failures actionable: distinguish request permission
  from the model allowlist, and explain authentication, model access, billing,
  IP restrictions and unsupported-region errors when the response identifies
  them. Bound response handling and display fixed safe guidance without echoing
  API keys or raw provider messages.

Direct upgrades from public 1.0.0, 1.1.0 and 1.1.1 preserve memories, data
volumes, credentials and embedding configuration. This patch adds no database
migration or embedding-model change.

To update an existing installation, run from its repository directory:

npm run update-persistent-memory -- --release 1.1.2

Windows PowerShell: npm.cmd run update-persistent-memory -- --release 1.1.2.
For a new installation, follow the
v1.1.2 installation instructions.

Validation covers automated installer/resource/provider checks, production
installer builds and Chrome wizard checks with controlled responses. A physical
Mac installation and the reported account-specific OpenAI rejection have not
been verified as resolved by this patch.

Persistent Memory 1.1.1 — Updates from GitHub Releases

GitHub release v1.1.1

Service	Version	Change
dashboard / dashboard layer	1.1.1	Published-release notices, exact-version GitHub links, and matching terminal update commands.
api	1.0.1	Exposes published-release metadata and directs installation through the host coordinator.
update-runner / update-flow layer	1.0.1	Anonymous stable-release discovery and immutable tag validation; sidecar provides status rather than installing updates.
update-coordinator	1.0.1	Pins published commits for each upgrade step and preserves those pins during recovery.
docs	1.1.1	Published-release update instructions, release links, and release-note validation and extraction.
onboarding	1.1.0	Existing resource-aware installation and embedding-provider choices retained.
shared / schema	1.1.0	Existing embedding catalog and schema retained; no new database migration.
worker / database	1.0.0	Existing processing and durable storage retained.
mcp / mcp-runtime	1.0.0	Existing memory tools and stream protocol retained.
core-tools	1.0.0	Existing maintenance tools retained.
graph / graphiti service	1.0.0	Existing memory graph behavior retained.
memory-vector / evidence-files / security-dlp / DLP service	1.0.0	Existing retrieval, evidence and sensitive-data checks retained.
dashboard-gateway / docker-control	1.0.0	Existing local gateway and service controls retained.

Updates now follow published stable GitHub Releases, so unreleased work on
master is not advertised or installed by the default update flow.

• Resolve the latest published release or an explicitly requested version to
  its exact Git tag commit, including annotated tags. Check the package version,
  public release lineage and release history at that immutable commit.
  Reject drafts, prereleases, unpublished tags and mismatched metadata.
• Keep update checks anonymous, cached and bounded, with retry backoff for
  temporary network failures and GitHub rate limits. No GitHub key is required.
• Pin every required upgrade step to a published commit. Reject changed tags
  and modified release worktrees; recovery retains the original release pins.
  Explicit --dev and --branch commands remain available for developer testing.
• Accept handoff from earlier public-release launchers only after checking
  their target version against its exact published tag commit.
• Route installation through the host coordinator and its snapshot/recovery
  flow. The dashboard sidecar provides update status and terminal guidance;
  the old direct container update path cannot bypass the coordinator.
• Link each release card and release-note entry to its exact GitHub Release.
  Validate those links during release preparation and extract the matching
  release body for publication. Include the shared source manifest and release
  resolver in the dashboard, sidecar and coordinator build artifacts they require.

Direct upgrades from public 1.0.0 and 1.1.0 preserve saved credentials,
embedding configuration, memories and data volumes. This patch adds no schema
migration or embedding-model change.

To update an existing installation, run from its repository directory:

npm run update-persistent-memory -- --release 1.1.1

Windows PowerShell: npm.cmd run update-persistent-memory -- --release 1.1.1.
For a new installation, follow the
v1.1.1 installation instructions.

Verification includes automated release, host/coordinator, API and dashboard
checks, production image builds, Windows deployment checks and live published
release discovery. A physical macOS update was not exercised for this patch.

Persistent Memory 1.1.0 — Smarter setup for your machine

Persistent Memory now helps you choose an embedding setup your computer can support. Smaller machines can use OpenAI or Voyage embeddings while keeping the memory database, graph, and dashboard local.

What is new

• Resource-aware installation: compare actual RAM, free RAM, disk space, CPU, and Docker resources with minimum and recommended budgets. Setup blocks installation below the minimum.
• Models that fit: suggest a suitable local Ollama model or API embeddings; disable local models when even the smallest option cannot fit. An 8 GB M2 Mac can use API embeddings and API fact extraction when the remaining requirements pass.
• API embedding choices: OpenAI text-embedding-3-small (value default) or text-embedding-3-large; Voyage voyage-4, voyage-4-large, or voyage-4-lite. Voyage uses its own key. Connection tests check provider access and vector dimensions.
• Safer retries: check image runtimes and compiled modules, build serially, and attempt one targeted rebuild/pull after an image check fails. Disk-full and storage I/O failures stop with guidance.
• Careful cleanup: remove unused artifacts with proven installation ownership; preserve persistent data, images used by other containers, and shared artifacts whose ownership is unknown.
• Protected settings: preserve saved keys and existing embedding configurations; allow a failed first-install configuration change only when the database is proven empty.
• Concurrent-operation fixes: prevent separate wizard tabs from changing settings or loading models during installation and resource tests.
• Clearer documentation: requirements tables, Windows browser GPU guidance, and explanations of GitHub Releases and Packages.

Install or update

Read the requirements and model guide and installation instructions. API embedding requests send memory/search text to the selected provider.

Existing public 1.0.0 installations can use the supported updater from their checkout:

npm run update-persistent-memory -- --branch master

In Windows PowerShell, use npm.cmd. The updater takes recovery snapshots and preserves existing credentials, memories, volumes, and embedding configuration. Changing an existing corpus to another embedding model requires a separate re-embedding migration. Update checks remain automatic; installing an update remains explicit.

Validation and limits

Root typechecks and test suites passed, together with 426 installer tests, 202 dashboard tests, 17 isolated desktop/narrow-browser wizard checks, and release/documentation validation. Independent reviews covered resource safety, concurrency, upgrade metadata, and the publication payload.

Physical macOS installation, real paid-provider connections, and deliberate Docker-storage corruption were not tested for this release. Resource budgets are planning estimates. Image rebuilding cannot repair a damaged Docker data disk or database. The new image preparation/repair flow applies to installation and retries; upgrades retain the existing updater recovery path.

This is a source release: ZIP and tar.gz downloads are available below. The installer builds Docker images locally. No npm package, prebuilt GHCR images, or standalone Windows/macOS executable is included.

Full release history · Changes since v1.0.0 · Release PR #3

Persistent Memory 1.0.0 — First public release

Give Claude, Codex, and MCP-compatible agents a memory that lasts beyond the chat.

This release preserves the original public 1.0.0 source merged on September 6, 2026, at commit ce7248a0e8025354e8294af85ed489238a4549c3. It is published retrospectively without changing that snapshot.

Highlights

• Personal memory services and databases running locally in Docker Linux containers.
• Guided Windows and macOS installation, including Ollama setup and Claude/Codex MCP and memory-rule registration.
• Semantic recall, document evidence, temporal relationships, and an interactive memory graph.
• Configurable API fact extraction and local Ollama embeddings.
• Optional Shared Memories connections from the personal dashboard.
• Automatic public GitHub update checks, with explicit update installation and recovery snapshots.
• Empty personal bootstrap without demo teams or sample access grants.

Installation

Follow the 1.0.0 README for prerequisites and the guided installer. GitHub supplies source ZIP and tar.gz downloads below. This release does not include a standalone executable, npm distribution, or prebuilt container package.

Known installation limitations

This historical version predates resource-aware model selection and installer API-embedding choices. Its recommended 4B local model can overwhelm an 8 GB Mac; use the newer release for smaller machines. Later fixes also improve disk checks, failed-image recovery, and cleanup.

Complete 1.0.0 release notes