Skip to content

Releases

Eunomia 2.0.4

Choose a tag to compare

@mars-s mars-s released this 16 Oct 07:16
f2ac7c0

Fix: saving a connector failed with "Internal server error" (#102)

On installs upgraded to 2.0.3, saving credentials for any connector (Notion, GitHub, Linear and the rest) could fail with internal. Connector rows created before 2.0.3 were missing a routing value that 2.0.3 made required. Saving now fills it in. No action needed: save again after updating.

Your agent can make its own entity kinds (#101)

Besides person, organisation, location, repository, file and symbol, an agent can now file entities under kinds it names itself, for example table, decision or defect. It creates a kind the first time it writes memory_write with a new subject_kind.

  • New MCP tool entity_types lists the built-in and custom kinds in your vaults.
  • Custom kinds show in search, recall, the graph and the Entities page filter.
  • Names are lowercase (letters, digits, _), up to 32 characters, up to 50 custom kinds per vault.

Connect a connector more than once (#100)

Two Gmail accounts, two Pocket devices: each is its own connection with its own name, credentials, sync, health and routing. On the Connectors page, use Add another connection under a connected kind. Existing connections keep working as the first connection of their kind.

Documents belong to a vault (#98)

  • Each vault's documents are stored separately.
  • Members of a shared vault can read and add its documents. Upload and list take an optional vault_id (API and MCP tools).
  • Cloning or merging a vault brings its documents along. Deleting a vault removes its documents.
  • Documents uploaded before 2.0.4 stay where they are and still download.

Page through a connector's records (#103)

A connector's page now has Previous and Next under its records, so you can see all of them, not just the latest 50. Search results page too. The search box on that page matches words (keyword search).

Upgrading

Update as usual. Three database migrations run at startup. For a moment, requests may answer "being upgraded, retry shortly".

Eunomia 2.0.3

Choose a tag to compare

@Qyrhal Qyrhal released this 11 Oct 17:14
7eebb38

Route connector data to vaults (#93)

Each connector now has a destination vault and optional rules (on each connector's setup page, under Connectors). For example, Pocket calls tagged work go to a team vault and everything else goes to your personal vault. Members of a vault see the records filed there; nobody else does.

  • Move existing data re-files what a connector already synced under its current routing. It is safe to run again: nothing is moved twice and facts aren't duplicated.
  • After merging vaults, repoint moves connectors that pointed at the old vault.
  • recall, reflect, search and entities_search take vault_ids to read from several vaults in one call.

Routing applies to records synced from now on. Data synced before 2.0.3 stays in your personal vault until you choose Move existing data.

Your AI knows how to use its memory, whatever it runs in

The memory skill (when to recall, what to save, a short summary before finishing a task) now comes from the MCP server itself. It arrives as server instructions, a prompt and a resource, and in the recall and memory_write tool descriptions, so it works even where hooks can't be installed (for example, Claude Code's auto mode). Hooks in connect-agents.sh are now optional and fail quietly.

Onboarding: "Connect your AI"

The setup wizard and dashboard have step-by-step guides for Claude Code, Codex, ChatGPT and other MCP clients. Each guide creates its own token and fills in your URL and token. See docs/agents.md.

Rename API tokens

Settings → API tokens: click a token's name to rename it.

Faster under load

Built for hundreds of agents connected at once. On a 250-agent MCP load test: about 570 req/s, p99 under 1 s, no errors.

  • Verified credentials are cached for 20 s. Revoking a token or session takes effect at once on the server that handled the revoke, and within 20 s on other replicas.
  • entities_search pages inside the database instead of loading every entity.
  • memory_write no longer waits for the model's contradiction check: it runs in the background, so superseded is no longer part of the result.
  • 8 database connections behind a first-come-first-served queue (EUNOMIA_DB_CONNECTIONS, EUNOMIA_DB_INFLIGHT).

Updater

  • Slow first start: after an update, the updater retries docker compose up for up to 30 minutes when a slow first start outlasts compose's health wait, and shows progress meanwhile.
  • Stuck services: if an update left a service stuck (for example the frontend in Created), the next run starts it. Services you stop by hand are left alone.

Includes everything in 2.0.2.

v2.0.2

Choose a tag to compare

@Qyrhal Qyrhal released this 10 Oct 13:01
9590476

Fix: the 1.x → 2.0 move into an org could stop on installs with consolidated observations

On 1.3+ installs, the one-time move into the org database could stop with memory: N in the old database, N-1 in the new. Nothing was lost (the move verifies before it finishes), but the app stayed offline.

Cause: observations from 1.3+ carry obs_subject, a field 2.0's stricter schema no longer has. SurrealDB 3 refuses such a row, and the copy skipped it.

Fix: rows are copied without fields the 2.0 tables don't define, and the move logs which fields it left out.

If you hit this on 2.0.0 or 2.0.1: update to 2.0.2 (or run docker compose up -d after updating). The move resumes and completes; your old data is untouched.

Includes 2.0.1's backup fix and everything in 2.0.0.

v2.0.1

Choose a tag to compare

@Qyrhal Qyrhal released this 10 Oct 09:14
8315ffd

Fix: updating from 1.x to 2.0 could stop at the pre-upgrade backup

On installs whose memories contain multi-line text (transcripts, for example), the update stopped with:

export of eunomia does not hold the rows the server counted

The data was never touched, but the app stayed offline until rolled back.

Cause: SurrealDB 2.3 exports keep raw newlines inside strings, and 2.0.0's new row-count check only read lines that start with INSERT.

Fix:

  • The counter now reads the whole export properly. It's exact on a real affected install and about 150× faster.
  • If a move ever fails, the message is now honest: the app is down, your data is untouched, and it gives the commands to retry or roll back.

If you hit this on 2.0.0: click Update now again, or run docker compose up -d. Your data is still on the original volume and moves on the next try.

Everything from 2.0.0 applies.

v2.0.0

Choose a tag to compare

@Qyrhal Qyrhal released this 10 Oct 07:33
da3fdad

Eunomia 2.0

A new foundation: SurrealDB 3.3, a database per org, background jobs, stronger auth, and uploaded documents. Update from Settings → Updates. Your data is moved and verified automatically, and the old data is kept for rollback.

Upgrading

  • Automatic database upgrade. On update, SurrealDB 2.x data is backed up (encrypted), copied, imported into 3.3 and verified table by table. The old volume is never modified, so rollback is always possible. This happens on docker compose up itself, so it works even if you skip 1.4.2 (recommended but not required).
  • One-time move into an org. A 1.x install becomes one org with its own database, and every existing user becomes a member. Record ids, logins and agent tokens keep working.
  • Installer 1.8.0 installs 2.0 fresh, or upgrades an existing install when re-run.

New

  • Documents. Upload text, Markdown, JSON, CSV and text-based PDFs (UI or MCP document_* tools). They're stored in SurrealDB's built-in file buckets (optionally on S3), split into passages, embedded, added to the entity graph, and downloadable as the original file. Backups include them.
  • A database per org, with its own database user. The server runs hardened, with an explicit function allow-list.
  • Background job queue with leases, retries and a scheduler leader, replacing the old sync loop.
  • Auth: scoped, vault-restricted and expiring tokens; MCP OAuth 2.1; signup modes; admin-only updates and HTTPS.
  • Debuggable by agents: stable error codes with trace ids, failure capsules and replay, JSON logs and optional OpenTelemetry.
  • Nightly encrypted backups that are refused unless per-table row counts match, plus restore drills.
  • Memory improvements from 1.4.1 are carried over:
    • org vaults are searched by default;
    • outdated facts are marked superseded;
    • entities can be found by the text of their facts, and possible duplicates are flagged;
    • Pocket calls feed the entity graph.

Fixed in review

  • The 1.x→2.0 move now keeps stored Pocket recordings.
  • The Pocket migration runs on the hardened server.
  • A failed upgrade's rollback never points the app at the temporary copy.
  • A memory edit and its stale summary are saved atomically.
  • Link-only connector changes are no longer dropped.
  • Vault clone and merge are all-or-nothing.
  • The migration lock can't be stolen mid-run.
  • The server refuses an empty ENCRYPTION_KEY, and replaces any database password that was derived from one.
  • Exactly-once job claims are tested across processes.
  • Re-uploading a document never shows both versions.

v1.4.2

Choose a tag to compare

@Qyrhal Qyrhal released this 10 Oct 04:24

Bridge to Eunomia 2.0

A maintenance release that prepares installs for 2.0. Eunomia 2.0 moves to SurrealDB 3.3 and gives each org its own database. There are no feature changes; take it with Settings → Updates as usual.

  • Updater: the updater now carries the SurrealDB 3 upgrade step. When you later update to 2.0, it takes an encrypted backup, copies the data to a new volume, upgrades and verifies it, and keeps your old volume untouched for rollback.
  • Backup service: a new backup service writes nightly encrypted backups (key: BACKUP_ENCRYPTION_KEY, generated for you).
  • SurrealDB stays on 2.3 in this release.

v1.4.1

Choose a tag to compare

@Qyrhal Qyrhal released this 09 Oct 16:42

Retrieval fixes from real agent use

  • Org vaults are no longer invisible (#83). Without vault_id, recall, reflect and entities_search search your personal vault plus every org vault you belong to. Each hit says which vault it came from. Writes still go to your personal vault by default.
  • Outdated facts drop out (#84). When new facts arrive, the chat model checks them against an entity's earlier facts. This happens on memory_write and after every sync. A fact a later one contradicts is marked superseded and leaves recall. Editing it makes it current again. Recency also counts for more, every hit carries its date, and reflect prefers newer memories.
  • Readable scores (#85). Recall scores run from 0 to 1, relative to the best hit.
  • Finding and de-duplicating entities (#86).
    • entities_search also matches the text of an entity's facts, so a handle that only appears in a fact still finds its entity.
    • A decorated name like Jane Doe (Acme/JD) now resolves to the existing Jane Doe instead of creating a duplicate, with Acme and JD added as aliases.
    • memory_write returns possible_duplicates when it creates an entity similar to an existing one.
  • Pocket calls feed the graph (#82). Extraction reads transcripts as conversations: the speakers and the people, organisations and places they mention, with dated facts. It doesn't create an entity for the meeting itself, and placeholder speakers like "Unknown" never become people.

Update from Settings → Updates, or re-run the installer.

v1.4.0

Choose a tag to compare

@Qyrhal Qyrhal released this 09 Oct 07:23

Pocket recordings, stored whole

  • Everything Pocket returns is kept. Each recording is stored verbatim in a new pocket_recording table. That covers the full speaker-labelled transcript, summary, action items, speakers, tags and the raw API response. get on any Pocket record returns the whole recording.
  • Searchable and in the memory graph. Each recording becomes a summary record plus transcript chunks small enough to embed. People, organisations and their relations are extracted from every chunk.
  • Fixed: summaries were never imported, and large syncs were capped by an over-limit page size. Existing Pocket connections refetch the last 30 days once, on their next sync.

Choose the chat model

The model that extracts entities and relations, synthesizes answers and powers chat can be set in two places:

  • OPENAI_CHAT_MODEL on the server;
  • Settings → OpenAI → Chat model, per user.

When it's left blank, Eunomia uses gpt-4o-mini on OpenAI. On any other endpoint it uses the first chat model the server lists, for example a local Ollama or vLLM model.

Delete a connector's data

Delete all data on a connector's page asks you to confirm twice. It then removes everything synced from that connector: records, search index, stored recordings, and the facts and relations extracted from them. The connection stays. The API is DELETE /api/connectors/{kind}/data.

Reliability

Syncs running at the same time no longer fail records on SurrealDB write conflicts. Conflicting writes are retried.

Update from Settings → Updates, or re-run the installer.

v1.3.1

Choose a tag to compare

@Qyrhal Qyrhal released this 09 Oct 00:45

Memory correctness fixes

  • Deleted records stay deleted (#64). They're excluded from get (REST and MCP), links and recall's graph arm. Facts learned from a record are kept, but no longer point back to it.
  • Beliefs track their facts (#65). Editing or deleting a fact marks the entity's observation stale, and the next consolidation rebuilds it from the facts that remain. Stale observations are no longer presented as current.
  • Recall is bounded (#68):
    • query length, limit, max_tokens and date ranges are validated;
    • the retrieval arms run in parallel, each with a deadline, so a slow embedding provider can't hold up keyword and graph results;
    • entity matching uses an index instead of scanning every entity.
  • Search filters are applied first (#69):
    • source, type and date filters apply before results are cut to the limit, so matches aren't lost;
    • paging and has_more are exact, and dates compare as datetimes;
    • body text has its own full-text index.
  • Merges are atomic (#72):
    • entity merges run in one transaction;
    • merging two entities that both have observations leaves one observation, to be rebuilt;
    • concurrent writes produce one entity and one observation per subject.

All verified by the full browser suite in CI against a real SurrealDB. Images are multi-arch (amd64 + arm64).

Updating

Settings → Updates → Update now.

v1.3.0

Choose a tag to compare

@Qyrhal Qyrhal released this 08 Oct 06:52

Highlights

  • HTTPS with Let's Encrypt. Turn it on in Settings → HTTPS or at install time with --domain and --acme-email. A built-in Caddy gets and renews the certificate. Updates keep it running and pick up Caddyfile changes.
  • Connectors that really sync. All 12 connectors in the UI pull from the provider's real API (Up Bank, PocketAI, GitHub, Slack, Notion, Linear, Gmail, Google Calendar, Discord, Spotify, Todoist, Stripe). Each one is tested against a mock of its API. The fake "Open Connector" is gone. Setup pages say where to get each token. New Connectors doc.
  • Vaults for anything. Every tool's vault_id also accepts a vault's name (vault_id: "Acme"). Use vaults for projects, services, clients, teams or a homelab. The memory skill tells agents to keep each area separate.
  • Graph back in 2D; the vector cloud stays 3D. The update badge in the sidebar now opens Settings → Updates from any page.

Security and correctness fixes

  • Vault isolation: 7 leaks closed, with an attack test suite. Two of them:
    • consolidating another user's entity could send their facts to your model endpoint;
    • pending invitees could export your vault.
  • The server's OpenAI key is never sent to a user-chosen model URL (#60). Redirects are not followed.
  • ENCRYPTION_KEY is now required (#62). Ciphertext is versioned, and data that fails to decrypt is never sent anywhere. Update now and the installer generate a key for installs that never set one; credentials saved without a key must be entered again.
  • Embeddings are tied to their provider and model (#70). The selected model is used, and malformed provider responses are rejected.
  • The in-app chat can no longer run destructive or sharing tools from a model's tool call (#73). Retrieved text is treated as untrusted data.
  • Sync reliability:
    • scheduled syncs, manual syncs and webhooks share one pipeline (#63);
    • sync cursors don't advance past failed writes (#66);
    • one scheduler, with no overlapping syncs (#67);
    • pagination and Gmail dates fixed (#71);
    • webhook bodies are capped at 1 MiB and slow uploads time out (#76).
  • Per-user connector base_url overrides are off unless the operator sets EUNOMIA_ALLOW_CONNECTOR_BASE_URL=1.

CI

The whole browser suite now runs in CI against a real SurrealDB (#74), including a fixed retrieval benchmark. Images are multi-arch (amd64 + arm64).

Updating

Settings → Updates → Update now. Installs older than v1.2.0: re-run the installer from the folder that contains your install.